Advances in

Orebody Modelling and

Strategic Mine Planning I
Old and New Dimensions in a Changing World
Editor: R Dimitrakopoulos

Number 17
 Advances in
Orebody Modelling and
Strategic Mine Planning I
Old and New Dimensions in a Changing World

Edited by

Roussos Dimitrakopoulos

The Australasian Institute of Mining and Metallurgy

Spectrum Series No 17

COSMO
Stochastic Mine Planning Laboratory

Published by
THE AUSTRALASIAN INSTITUTE OF MINING AND METALLURGY
Level 3, 15 - 31 Pelham Street, Carlton Victoria 3053 Australia
 © The Australasian Institute of Mining and Metallurgy 2010

All papers published in this volume were refereed prior to publication.

The Institute is not responsible as a body for the facts and opinions advanced in any of its publications.

ISBN 978 1 921522 21 5

Desktop published by:

Kylie McShane and Olivia Tet Fong for
The Australasian Institute of Mining and Metallurgy

Compiled on CD ROM by:

Visual Image Processing Pty Ltd
PO Box 3180
Doncaster East VIC 3109
 Reviewers
We would like to thank the following people for their contribution
towards enhancing the quality of papers included in this volume.

Enrique Acuna Ute Mueller

Jorg Benndorf Hussein Mustapha

Alexandre Boucher Mike O‘Brien

Marcus Brazil Harry Parker

Peter Carter Richard Peattie

Snehamoy Chatterjee Salih Ramazan

Michel Dagbert Jean-Michel Rendu

Roussos Dimitrakopoulos Andrew Richmond

Tamer Dincer Sabry Abdel Sabour

Peter Dowd Michael Samis

Marcelo Godoy Peter Stone

Nicole Grieco Olivier Tavchandjian

John Hadjigeorgiou Erkan Topal

Geoff Hall Marcus Volz

Xu Huan Chris Wharton

Arja Jewbali David Whittle

Brett King Jeff Whittle

Amina Lamghari Gavin Yeates

Conor Meagher
 Foreword
In an ever-changing world of uncertain markets, mismatches between demand and reserve base growth
(supply), the increasingly critical issue of sustainable development, emerging technologies and new
technical problems and their solutions, this book represents a contribution to both old and new
dimensions of a most intricate, demanding and fundamentally important technical aspect of mining
ventures and the industry. These and other issues add complexity and uncertainty to the already
challenging domains of orebody modelling and strategic mine planning, which are arguably the
backbone of our industry.

It is a privilege to introduce the new Spectrum Series Volume 17: Advances in Orebody Modelling and
Strategic Mine Planning I, focusing on ‘Old and new dimensions in a changing world’. This
publication presents the latest advances, technologies, practices and concerns in strategic mine planning
and orebody modelling, thus updating the developments and applications reported in the remarkably
successful 2005 Spectrum Series 14 and its second revised edition in 2007.

The unique qualities of this volume are as evident today as they were in the previous Spectrum Series
volume, thanks to many factors. To begin with, this volume has an impressive author list of
international experts, including a new and promising generation of young colleagues in our field who
we hope to see continuing our efforts in the decades to come; these contributions by so many respected
colleagues is reassuring for our future success. Second, the ongoing support and involvement of the
global mining industry has become indispensable; and finally, the exceptional and continued
collaboration of the major national institutes who have supported this volume and the related
international symposium, held 16 - 18 March 2009, in Perth, Western Australia: The AusIMM
(Australia), CIM (Canada), SME (USA) and The SAIMM (South Africa). In the present global mining
environment, I believe this is a substantial effort by these institutes to enhance professional excellence
in the critical field addressed by this volume. I am proud that this collaboration has been fostered by all
involved, to the benefit of our profession.

This volume comprises 41 papers under the following topics:

• Modelling and Planning in the New World
The volume opens with papers examining global optimisation, mining optimisation and management
of multiple objectives, risk quantification, joint effects of metal and market uncertainties in mine
design, lessons learned and old and new paradigms.
• Modelling Orebodies and Applications
This section stresses new developments for modelling orebodies in different ways, including complex
geology, indirect data and optimising drilling, while keeping in mind the needs of the industry for
methods and implementations facilitating their routine use in the industry environment.
• Mining Optimisation and Case Studies
This group of papers complements the previous section, stressing case studies involving different
commodities.
• Large-Scale Integrated Applications
This section presents major case studies that use new technologies and new approaches, along with
their applications, which include open pit mine design, production scheduling under uncertainty,
cut-off grade optimisation, modelling the supply chain and assessing deep sited gold reserv es.
• Uncertainty Models and Recoverable Reserves
This section is focused on the related issues underpinning the sustainability of the mining industry
and includes new, state-of-the-art simulation approaches and case studies.
• Optimisation is Moving Underground
This section reflects a rapid growth in the application of optimisation approaches in various aspects
of underground mining, from access to ventilation as well as production scheduling.
• The New World and Strategic Innovation
This final section deals with the changing world. It stresses major applications and frameworks for
stochastic mine planning optimisation, and extends to optimising under the joint uncertainty of metal,
markets and mining, as well as newest state-of-the-art high-order simulation approaches and
application of simulation.

As always, a volume such as this is due to the combined efforts of several colleagues over several years.
I would particularly like to thank our colleagues and international experts: Jeff Whittle, Gavin Yeates,
Wynand Kleingeld, Jean-Michel Rendu, Rick Allan, Ian Douglas, Vaughan Chamberlain, Edson
Ribeiro, Brian Baird, David Whittle, Malcolm Thurston, Olivier Tavchandjian, Mike O’Brien,
Jean-Yves Cloutier, Peter Ravenscroft, Jaimie Donovan, Peter Stone, Malcolm Thurston, Andre
Journel, Christophe Pierre and our colleagues involved with the COSMO Lab. Their support over the
past several years has been invaluable. I would further like to acknowledge the diligent work of the
reviewers of the papers in this volume who are listed on page iii, and thank the authors for the high
quality of their contributions.

Once again, support from the industry has been remarkable, with an impressive list of major sponsors
whose long-term multifaceted support I gratefully acknowledge: AngloGold Ashanti, Barrick, BHP
Billiton, De Beers, GEMCOM Australia, Newmont and Vale.

I believe the underlying reason for the unprecedented support for this and previous volumes on strategic
mine planning is the critical importance to the mining industry of new leading-edge technologies that
quantify uncertainty and manage risk in decision-making on a global scale. This uncertainty stems from
both the uncertain supply of metal and an uncertain future market demand. Given the high risk/high
reward nature of mining and the current global economic downturn, this volume of proceedings (i)
contributes new technologies that add substantial value to mining assets, and (ii) assists with the related
knowledge dissemination in a rapidly changing world.

It takes an enormous effort to produce a Spectrum Volume like the one in your hands. I would like to
thank The AusIMM and its staff for its continuing support for our area of work, including Miriam Way,
who has been the source of encouragement and leadership behind several professional development
activities over the past decade. Thanks are also in order to Katy Andrews, Kylie McShane, Kristy
Pocock and Alison McKenzie for their tremendous effort (under conditions of editorial uncertainty)
towards this Spectrum Volume and its quality.

On a personal note, this year marks a decade of collaboration with The AusIMM; I would like to
express my gratitude to this remarkable organisation for its contribution to the mining industry and our
profession. Both Australia and the global mining community are lucky to have such an Institute. Thank
you AusIMM!

Roussos Dimitrakopoulos
COSMO – Stochastic Mine Planning Laboratory
McGill University, Montreal
An initiative of

COSMO
Stochastic Mine Planning Laboratory

Sponsors
 Contents
Modelling and Planning in the New World

The Global Optimiser Works – What Next? J Whittle 3

Optimal Mining Principles B King 7

Stochastic Mine Planning – Methods, Examples and Value in an R Dimitrakopoulos 13

Uncertain World

A Risk Analysis Based Framework for Strategic Mine Planning M Godoy 21

and Design – Method and Application

Accounting for Joint Ore Supply, Metal Price and Exchange Rate S A Abdel Sabour and 29
Uncertainties in Mine Design R Dimitrakopoulos

Production Scheduling Under Metal Uncertainty – Application of A Leite and R Dimitrakopoulos 35

Stochastic Mathematical Programming at an Open Pit Copper
Mine and Comparison to Conventional Scheduling

Modelling Orebodies and Applications

Regional Three-Dimensional Modelling of Iron Ore Exploration V Osterholt, O Herod and 43

Targets H Arvidson

Conditional Simulation of Grades Controlled by Geological C Xu and P A Dowd 51

Indicators

Mining Geostatistics with the Open Source SGeMS Software and A Boucher 59
the GsTL Library

Conditional Simulation Methods to Determine Optimum Drill Hole C Boyle 67

Spacing

Transformation Methods for Multivariate Geostatistical Simulation E M Bandarian, U A Mueller, 79

– Minimum/Maximum Autocorrelation Factors and Alternating J Fereira and S Richardson
Columns Diagonal Centres

Conditional Wavelet Based Simulation of Non-Stationary E Gloaguen and 91

Geologic Fields Using Geophysical and Model Analogue R Dimitrakopoulos
Information
 Mining Optimisation and Case Studies

Optimal Life-of-Mine Scheduling for a Bauxite Mine M Zuckerberg, J van der Riet, 101
W Malajczuk and P Stone

Coal Mining in the Hunter Valley – A Strategic Mine Planning C Wharton 107
Case Study

A New Methodology for Flexible Mine Design B Groeneveld, E Topal and 113
B Leenders

Pit Optimisation and Mine Scheduling at the Montes Claros De B Y Tomaselli, E A Silva, 123
Goiás Nickel Laterite Deposit, Brazil G J de Carvalho Pereira,
J A Hilário, M C Botelho and
P C Rodriguez

Technogenic Iron Ore Formations in Siberia as a Secondary P A Filippov, S A Neverov and 129
Resource for Production of Ferrous, Rare, Non-Ferrous and A A Neverov
Noble Metals

Advancing Strategic Mine Planning – A Software Perspective C Tonkin 135

Large-Scale Integrated Applications

Integrated Strategic Planning at EKATI Diamond Mine B Coutts 143

Stochastic Long-Term Production Scheduling of Iron Ore J Benndorf and 151

Deposits – Integrating Joint Multi-Element Geological Uncertainty R Dimitrakopoulos

Overall Optimisation of Yerilla Project T Dincer and K Reynolds 159

Investigating the Variability of Key Coal Quality Parameters in J Benndorf 167

Continuous Mining Operations when Using Stockpiles

Cut-Off Grade Estimation – Old Principles Revisited – Application J-M Rendu 173
to Optimisation of Net Present Value and Internal Rate of Return

Using Grade Uncertainty to Quantify Risk in the Ultimate Pit S Robins 179
Design for the Sadiola Deep Sulfide Prefeasibility Project, Mali,
West Africa
 Uncertainty Models and Recoverable Reserves

Angle Reverse Circulation Versus Blastholes in Ore Control – I Douglas, B Perry, W Hardtke 193
Sampling Quality Versus Data Density – A Case Study and L Allen

Impact of Bench Height, Cut-Off Grade and Production Spacing W Hardtke, L Allen, B Perry 199
on Misclassification in Ore Control – Challenges Produced by and
High Nugget, Low Continuity Ores I Douglas

A Geology-Based Joint Conditional Simulation of a Lateritic J A Lopes, C F Rosas, 205

Nickel Deposit – Quantification of Risk in Grade Tonnage Curves J B Fernandes and G A
and Resource Categorisation Vanzela

Resource/Reserve Classification with Integrated Geometric and R Dimitrakopoulos, M Godoy 215

Local Grade Variability Measures and C L Chou

Reducing the Geological Risk in Mining an Orebody by Using C Kemp, P Du Pisani, A Bray 223
Borehole Radar for Strategic Mine Planning and G Chitiyo

Optimisation is Moving Underground

Strategic Optimisation of a Vertical Hoisting Shaft in the Callie M G Volz, M Brazil and 235
Underground Mine D A Thomas

Optimising Value for an Underground Project Configuration T Elkington, R Durham and 241
P Myers

Classification of Mining Methods for Deep Orebodies V Oparin, A Tapsiev and 249
A Freidin

Resource Risk Quantification for a Deep Level Gold Mining T Flitton and R Peattie 257
Operation

An Application of Genetic Algorithms to the Main Ventilation E Acuña, S Hardcastle, L Fava, 263
System Optimisation for Multi-Level Mine Operations – The P Dunn and S Hall
Impact of the Fitness Function

The New World and Strategic Innovation

Applying Sequence Optimisation Technology to Longwall Coal L Rocchi, P Carter and P Stone 271
Mining

Combining Optimisation and Simulation to Model a Supply Chain P Bodon, C Fricke, T 277
from Pit to Port Sandeman and C Stanford
Direct Net Present Value Open Pit Optimisation with Probabilistic A Richmond 285
Models

Pushback Design of Open Pit Mines Under Geological and C Meagher, S A Abdel Sabour 291
Market Uncertainties and R Dimitrakopoulos

Faulted Geological Model Simulation of the Resolution Porphyry G Verly, K Brisebois, W Hart 299
Copper Deposit and J Hammitt

Geologically Enhanced Simulation of Complex Mineral Deposits H Mustapha and 309

Through High-Order Spatial Cumulants R Dimitrakopoulos

Stochastic Mine Planning – Example and Value from Integrating A Jewbali and R 321
Long- and Short-Term Mine Planning Through Simulated Grade Dimitrakopoulos
Control, Sunrise Dam, Western Australia

Author Index 329

 MONOGRAPH SERIES
1 • Detrital Heavy Minerals in Natural Accumulates George Baker 1962
2. • Research in Chemical and Extraction Metallurgy Ed: J T Woodcock, 1967
A E Jenkins
and G M Willis
3. • Broken Hill Mines - 1968 Ed: M Kadmanovich 1968
and J T Woodcock
4. • Economic Geology of New Zealand Ed: G J Williams 1974
5. • Economic Geology of Australia and Papua New Guinea - 1 Metals Ed: C L Knight 1975
6 • Economic Geology of Australia and Papua New Guinea - 2 Coal Ed: D M Traves 1975
and D King
7. • Economic Geology of Australia and Papua New Guinea - 3 Petroleum Ed: R B Leslie 1976
H J Evans
and C L Knight
8. • Economic Geology of Australia and Papua New Guinea - 4 Industrial Ed: C L Knight 1976
Minerals and Rocks
9. Field Geologists’ Manual
1st Edition Ed: D A Berkman 1976
2nd Edition and W Ryall 1982
3rd Edition Ed: D A Berkman 1989
4th Edition Ed: D A Berkman 2001
10 • Mining and Metallurgical Practices in Australasia Ed: J T Woodcock 1980
(the Sir Maurice Mawby Memorial Volume)
11. • Victoria’s Brown Coal - A Huge Fortune in Chancery Ed: J T Woodcock 1984
(the Sir Willis Connolly Memorial Volume)
12. • Australasian Coal Mining Practice
1st Edition Ed: C H Martin 1986
2nd Edition Ed: C H Martin 1993
and A J Hargraves
3rd Edition Ed: R J Kininmonth 2009
and E Y Baafi
13. • Mineral Deposits of New Zealand Ed: Dr D Kear 1989
14. Geology of the Mineral Deposits of Australia and Papua New Guinea Ed: F E Hughes 1990
15. The Rocks Speak H King 1989
16. • Hidden Gold - The Central Norseman Story J D Campbell 1990
17. • Geological Aspects of the Discovery of Some Important K R Glasson 1990
Mineral Deposits in Australia and J H Rattigan
18. • Down Under - Mineral Heritage in Australasia Sir Arvi Parbo 1992
19. Australasian Mining and Metallurgy Ed: J T Woodcock 1993
(the Sir Maurice Mawby Memorial Volume) and K Hamilton
20. • Cost Estimation Handbook for the Australian Mining Industry Ed: M Noakes 1993
and T Lanz
21. • History of Coal Mining in Australia Ed: A J Hargraves, 1993
(the Con Martin Memorial Volume) R J Kininmonth,
C H Martin
and S M C Saywell
22. Geology of Australian and Papua New Guinean Mineral Deposits Ed: D Berkman 1998
and D Mackenzie

Copies of all books currently in print can be obtained from

The Institute office - Tel (03) 9662 3166

Key: • Out of print

23. Mineral Resource and Ore Reserve Estimation — Ed: A C Edwards 2001
The AusIMM Guide to Good Practice
24. Australian Mineral Economics Ed: P Maxwell and P Guj 2006
25. Geology and Exploration of New Zealand Mineral Deposits Ed: A B Christie and 2006
R L Brathwaite

Copies of all books currently in print can be obtained from

The Institute office - Tel (03) 9662 3166

Key: • Out of print

 PUBLICATIONS OF THE AUSTRALASIAN INSTITUTE
OF MINING AND METALLURGY

CONFERENCE, SYMPOSIUM AND MISCELLANEOUS

1972 1/72 SI * Project Evaluation and Management, Melbourne
2/72 * Annual Conference, Newcastle
1973 1/73 S2 * Subsidence, Illawarra
2/73 S3 * Mine Filling, North West Queensland
3/73 S4 * Transportation, Sydney
4/73 S8 * Mine Fires, Southern Queensland
5/73 * Annual Conference, Western Australia
1974 1/74 S5 * Support in Pillar Extraction, Illawarra
2/74 S6 * Recent Technical and Social Advances in the North Australian Minerals Industry,
North West Queensland
3/74 S7 * Pellets and Granules, Newcastle and District
4/74 * Annual Conference, Southern and Central Queensland
1975 1/75 S9 * People and the Mining Industry - The Future, Broken Hill
2/75 S 10 * Occupational Safety in Mines, Southern Queensland
3/75 S 11 * Australian Black Coal, lllawarra
4/75 * Annual Conference, South Australia
1976 1/76 S13 * Landscaping and Land Use Planning as Related to Mining Operations, Adelaide
2/76 * Design and Construction of Tunnels and Shafts, Melbourne - 2 nd Australian
Tunnelling Conference
3/76 S 14 * Thick Seam Mining by Underground Methods, Central Queensland
4/76 S 15 * Sampling Practices in the Mineral Industries, Melbourne
5/76 * Annual Conference, lllawara
1977 1 /77 S 16 * Apcom ‘77, Brisbane
2/77 S 18 * Coal Borehole Evaluation, Southern Queensland
3/77 S 17 * Underground Operators’ Conference, Broken Hill
4/77 * Annual Conference, Hobart
1978 1 /78 S 19 * Mill Operators’ Conference, North West Queensland
2/78 S20 * Rock Breaking Equipment and Techniques, Melbourne
3/78 * International Resource Management, Canberra
4/78 * Annual Conference, Townsville
1979 1/79 S21 * Utilisation of Steelplant Slags, lllawarra
2/79 S22 * Estimation and Statement of Mineral Reserves, Sydney
3/79 * Annual Conference, Perth
1980 1/80 * Annual Conference, New Zealand
2/80 S23 * Australia/Japan Extractive Metallurgy Symposium, Sydney
3/80 S/24 * Occurrence, Prediction and Control of Outbursts in Coal Mines, Southern
Queensland
4/80 S25 * Management in the Mining Industry, Melbourne

Copies of all books currently in print can be obtained from The Institute office
Tel (03) 9662 3166 or Fax (03) 9662 3662

* = Out of print
The ‘S’ numbers in the third column refer to an older identifying number for Symposia, the numbers preceeding the ‘S’ number signify the new
publication ordering number.
1981 I/81 S26 * International Blast Furnace Hearth and Raceway Symposium, Newcastle
2/81 * Fourth Australian Tunnelling Conference
3/81 S27 * Ignitions, Explosions and Fires, Wollongong
4/81 * Annual Conference, Sydney
5/81 S28 * Strip Mining 45 Metres and Beyond, Central Queensland
1982 1/82 S29 * Off Highway Truck Haulage Conference, Newman
2/82 S30 * Mill Operators’ Conference, North West Queensland
3/82 S31 * Underground Operators’ Conference, West Coast Tasmania
4/82 * Annual Conference, Melbourne
5/82 S32 * Carbon-ln-Pulp Technology for the Extraction of Gold, Perth and Kalgoorlie,
(Reprinted 1988)
6/82 S33 * Seam Gas Drainage with Particular Reference to the Working Seam, Wollongong
1983 1/83 S34 * Computers in Mining, Southern Queensland
2/83 * Annual Conference, Broken Hill
3/83 S35 * Project Development Symposium, Sydney
4/83 S37 * Ventilation of Coal Mines, Wollongong
5/83 S40 * Principles of Mineral Flotation (The Wark Symposium), Adelaide
1984 1/84 S36 * Metallurgy Symposium, Melbourne
2/84 S38 * Coal and Mineral Sizing, Wollongong
3/84 * Annual Conference, Darwin
4/84 S39 Gold Mining, Metallurgy and Geology, Kalgoorlie
1985 1/85 S41 * Smelting and Refining Operators’ Symposium, North Queensland
2/85 S42 * Underground Operators’ Conference, Kalgoorlie
3/85 * Annual Conference, Brisbane
4/85 S43 * Scientific and Technological Developments in Extractive Metallurgy (G K
Williams Memorial Volume), Melbourne
1986 1/86 * l3th Congress The Council of Mining and Metallurgical Institutions, Singapore, 6
Volumes
2/86 S44 * Selective, Open Pit Gold Mining Seminar, Perth
3/86 S45 * Ground Movement and Control Related to Coal Mining, Wollongong
4/86 S46 * Australia: A World Source of Illmenite, Rutile, Monazite and Zircon Conference,
Perth
5/86 S47 * Second Project Development Symposium, Sydney
6/86 S48 * Large Open Pit Mining Conference, Newman
7/86 S49 * Education and Research for the Mineral Industry for the Future, Melbourne
8/86 * The AuslMM 10 Year lndex
1987 1/87 * Vl Australian Tunnelling Conference: Bore or Blast, Melbourne
2/87 S50 * Risk and Survival Seminar, Canberra
3/87 * Annual Conference, Newcastle: Coal Power ‘87
4/87 S51 * Research and Development in Extractive Metallurgy, Adelaide
5/87 * Leslie Bradford Golden Jubilee Oration
6/87 S52 * Mining and Environment: A Professional Approach, Brisbane
7/87 * Pacrim ‘87, Gold Coast, Queensland
8/87 S53 * Dense Medium Operators’ Conference, Brisbane

Copies of all books currently in print can be obtained from The Institute office
Tel (03) 9662 3166 or Fax (03) 9662 3662

* = Out of print
The ‘S’ numbers in the third column refer to an older identifying number for Symposia, the numbers preceeding the ‘S’ number signify the new
publication ordering number.
 9/87 S54 * Equipment in the Minerals Industry: Exploration Mining and Processing
Conference, Kalgoorlie
10/87 S55 * Resources and Reserves, Sydney
11/87 * South Australia’s Mining Heritage
1988 1/88 S56 * 21st Century Higher Production Coal Mining Systems Symposium, Wollongong
2/88 * The Second International Conference on Prospecting in Arid Terrain, Perth
3/88 S57 * Third Mill Operators’ Conference, Cobar
4/88 S58 * Underground Operators’ Conference, Mount Isa
5/88 * Fourth International Mine Ventilation Congress, Brisbane (Proceedings and
Addendum volume)
6/88 * Annual Conference, Sydney: Minerals and Exploration at the Crossroads: The
International Outreach
7/88 S59 * Second AuslMM Mineral Heritage Seminar, Sydney
8/88 S60 * Economics and Practice of Heap Leaching in Gold Mining Workshop, Cairns
9/88 * Third International Mine Water Congress, Melbourne
10/88 S61 * Explosives in Mining Workshop, Melbourne
1989 1/89 * Mineralogy and Petrology, Sydney, February
2/89 * Second Large Open Pit Mining Conference, Latrobe Valley Vic
3/89 * NQ Gold ‘89 Conference, Townsville Qld
4/89 * Annual Conference, Perth-Kalgoorlie: Education, Training and Professional
Development; Industrial Minerals; Project Development/Processing
5/89 * Mineral Fuel Alternatives and the Greenhouse Effect, July 1989
6/89 Non-ferrous Smelting Symposium: 100 Years of Smelting and Refining
Operations in Port Pirie, SA September 1989
7/89 * Dewatering Technology and Practice Conference, Brisbane October 1989
8/89 MINVAL ‘89, Mining and Petroleum Valuation 1989, Sydney September 1989
1990 1/90 * Ore Reserve Estimates - The Impact on Miners and Financiers, Melbourne, March
1990
2/90 Annual Conference, The Mineral Industry in New Zealand, Rotorua New
Zealand, March 1990
3/90 Pacific Rim Congress, Gold Coast Qld, May 1990
4/90 * Mining Industry Capital and Operating Cost Estimation Conference, Sydney,
June 1990
5/90 * Third International Symposium on Rock Fragmentation by Blasting, Brisbane,
August 1990
6/90 * Sir Edgeworth David Memorial Oration, May 1990
7/90 * Mine Geologists’ Conference, Mount Isa, October 1990
1991 1/91 * Fourth Mill Operators’ Conference, Burnie Tas, March 1991
2/91 * World Gold ‘91, Cairns Qld, April 1991
3/91 Mining Industry Optimisation Conference, Sydney, June l991
4/91 * PNG Geology, Exploration and Mining Conference, Rabaul, June 1991
5/91 Qld Coal Symposium, Brisbane, August 1991
6/91 * Reliability Production and Control in Coal Mines, Wollongong, September 1991
7/91 * Fifth AuslMM Extractive Metallurgy Conference, Perth, October 1991

Copies of all books currently in print can be obtained from The Institute office
Tel (03) 9662 3166 or Fax (03) 9662 3662

* = Out of print
The ‘S’ numbers in the third column refer to an older identifying number for Symposia, the numbers preceeding the ‘S’ number signify the new
publication ordering number.
1992 1/92 * Enviromine Australia, Sydney NSW, March 1992
2/92 The AuslMM Annual Conference, ‘The State-of-the-Art - A Product of 100 Years
of Learning’, Broken Hill NSW, May 1992
3/92 ‘Energy, Economics and Environment’ Gippsland Basin Symposium, Melbourne,
June 1992
4/92 Arnold Black Mineral Heritage Oration
5/92 The Man from ASARCO: a life and times of Julius Kruttschnitt
6/92 5th Underground Operators’ Conference, Ballarat, July 1992
7/92 * 11th International Conference on Ground Control in Mining, Wollongong, July
1992
8/92 * Third Large Open Pit Mining Conference, Mackay, August 1992
9/92 * Extractive Metallurgy of Gold and Base Metals Conference, Kalgoorlie, October
1992
10/92 * Sampling Practices in the Minerals Industry, Mount Isa, November 1992
11/92 * Rehabilitate Victoria, Latrobe Valley, November 1992
1993 1/93 Mining People - A Century
2/93 The AuslMM Centenary Conference, Adelaide, March 1993
3/93 XVIII International Mineral Processing Congress, Sydney, May 1993
4/93 * Narrow Vein Mining Seminar, Bendigo, June 1993
5/93 * International Mining Geology Conference, Kalgoorlie, July 1993
6/93 Vlll Australian Tunnelling Conference, Sydney, August 1993
7/93 * World Zinc ‘93 - International Symposium, Hobart, October 1993
1994 1/94 1994 AuslMM Student Conference, Brisbane, April 1994
2/94 * PNG Geology, Exploration and Mining Conference, Lae, PNG, June 1994
3/94 No Two The Same by Bert Mason
4/94 Sixth Extractive Metallurgy Conference, Brisbane July 1994
5/94 * 1994 AuslMM Annual Conference, Darwin, August 1994
6/94 * 4th Large Open Pit Mining Conference, Perth, September 1994
7/94 * Recent Trends in Heap Leaching, Bendigo, September 1994
8/94 * Maintenance in the Mining and Metallurgical Industries,Wollongong, October
1994
9/94 * Fifth Mill Operators’ Conference, Roxby Downs, October 1994
10/94 Mineral Valuation Methodologies 1994, Sydney, October 1994
11/94 Victorian Mining Week Conference, Melbourne, October 1994
1995 1/95 1995 AuslMM Annual Conference, Newcastle, March 1995
2/95 Sir Maurice Mawby Memorial Oration
3/95 World’s Best Practice in Mining and Processing Conference, Sydney, May 1995
4/95 APCOM XXV 1995 Conference, Brisbane, July 1995
5/95 Mineral Valuation Methodologies 1994, Sydney, October 1994 (revised)
6/95 * EXPLO 95 Conference, Brisbane, September 1995
7/95 Underground Operators’ Conference, Kalgoorlie, November 1995
8/95 * Young Professionals’ Conference, Mt Isa, October 1995
9/95 * PACRIM ‘95 Congress, Auckland, New Zealand, November 1995
10/95 * Ethics, Liability and the Technical Expert, Sydney, December 1995

Copies of all books currently in print can be obtained from The Institute office
Tel (03) 9662 3166 or Fax (03) 9662 3662

* = Out of print
The ‘S’ numbers in the third column refer to an older identifying number for Symposia, the numbers preceeding the ‘S’ number signify the new
publication ordering number.
1996 1/96 * 1996 AusIMM Annual Conference, Perth, March 1996
1a/96 * 1996 AusIMM Annual Conference Supplementary Volume, Perth, March 1996
2/96 * Ethics, Liability and the Technical Expert, Sydney, March 1996
3/96 * Entrepreneurs and Partners, Sydney, July 1996
4/96 * Contract Operators’ Conference, Kalgoorlie, October 1996
5/96 Asia/Pacific Mining Communications Summit, Singapore, November 1996 -
Withdrawn
6/96 * Nickel ‘96, Kalgoorlie, November 1996
1997 1/97 1997 AusIMM Annual Conference, Ballarat, March 1997
2/97 * World Gold ‘97 Conference, Singapore, September 1997
3/97 Sixth Mill Operators’ Conference, Madang, PNG, October 1997
4/97 Gem 97, Madang, PNG, October 1997
5/97 * Contract Operators’ Conference, Brisbane, Qld, October 1997
6/97 Third International Mining Geology Conference, Launceston, Tas, November
1997
7/97 Mindev 97 - The International Conference on Mine Project Development,
Sydney, November 1997
8/97 1997 AusIMM Travelling Technology Forum, Singleton, NSW, March 1997
1998 1/98 * MINEFILL ‘98 - The Sixth International Symposium on Mining with Backfill,
Brisbane, Qld, April 1998
2/98 * AusIMM’98 - The Mining Cycle, Mount Isa, Qld, April 1998
3/98 * Seventh Underground Operators’ Conference, Townsville, Qld, June/July 1998
4/98 Mine to Mill Conference, Brisbane, Qld, October 1998
5/98 Third Regional APCOM - Computer Applications in the Minerals Industries
International Symposium, Kalgoorlie, WA, December 1998
6/98 COAL98 - First Australasian Coal Operators’ Conference, Wollongong, NSW,
February 1998
1999 1/99 10th Australian Tunnelling Conference, Melbourne, Vic, March 1999
1a/99 10th Australian Tunnelling Conference Keynote Addresses and Asia–Pacific
Forum, Melbourne, Vic, March 1999
2/99 Students and Young Professionals Conference, Perth, WA, July 1999 (N/A)
3/99 ICARISM ’99 Conference, Perth, WA, September 1999
4/99 PACRIM ’99 Congress, Bali, October 1999
5/99 EXPLO ’99 Conference, Kalgoorlie, WA, November 1999
2000 1/2000 Southern Africa - Australia Mineral Sector Synergies Symposium, Canberra,
ACT, March 2000
2/2000 After 2000 - The Future of Mining, Sydney, NSW, April 2000
3/2000 4th International Mining Geology Conference, Coolum, Qld, May 2000
4/2000 Young Leaders 2000, Sydney, NSW, July 2000
5/2000 MINPREX 2000, Melbourne, Vic, September 2000
6/2000 Seventh Mill Operators’ Conference, Kalgoorlie, WA, October 2000
7/2000 MassMin 2000, Brisbane, Qld, October - November 2000
2001 1/2001 Strategic Mine Planning 2001, Perth, WA, March 2001
2/2001 AusIMM Youth Congress 2001, Brisbane, Qld, May 2001
3/2001 International Heavy Minerals Conference, Fremantle, WA, June 2001
4/2001 EXPLO 2001, Hunter Valley, NSW, October 2001

Copies of all books currently in print can be obtained from The Institute office
Tel (03) 9662 3166 or Fax (03) 9662 3662

* = Out of print
The ‘S’ numbers in the third column refer to an older identifying number for Symposia, the numbers preceeding the ‘S’ number signify the new
publication ordering number.
 5/2001 Mineral Asset Valuation Issues 2001, Sydney, NSW, October 2001
6/2001 Coal2001, Coal Mining Operators’ Geotechnology Colloquium, Wollongong,
NSW, February 2001
2002 1/2002 Young Leaders 2002, Kalgoorlie, WA, March 2002
2/2002 Metallurgical Plant Design and Operating Strategies, Sydney, NSW, April 2002
3/2002 CMMI Congress, Cairns, Qld, May 2002
4/2002 Green Processing Conference, Cairns, Qld, May 2002
5/2002 Eighth Underground Operators’ Conference, Townsville, Qld, July 2002
6/2002 AusIMM 2002 Conference, 150 Years of Mining, Auckland, New Zealand,
September 2002
7/2002 Iron Ore Conference, Perth, WA, September 2002
8/2002 Value Tracking Symposium, Brisbane, Qld, October 2002
9/2002 Coal2002, Third Australasian Coal Operators’ Conference, Wollongong NSW,
February 2002
2003 1/2003 Twelfth International Symposium on Mine Planning and Equipment Selection,
Kalgoorlie, WA, April 2003
2/2003 Young Leaders 2003, Brisbane, Qld, April 2003
3/2003 Sixth International Conference Acid Rock Drainage, Cairns, Qld, July 2003
4/2003 Eighth Mill Operators’ Conference, Townsville, Qld, July 2003
5/2003 Mining Risk Management Conference, Sydney, NSW, September 2003
6/2003 Water in Mining 2003, Brisbane, Qld, October 2003
7/2003 Fifth Large Open Pit Conference, Kalgoorlie, WA, November 2003
8/2003 Fifth International Mining Geology Conference, Bendigo, Vic, November 2003
9/2003 Coal2003, Fourth Australasian Coal Operators’ Conference, Wollongong NSW,
February 2003
2004 1/2004 New Leaders 2004, Ballarat, Vic, April 2004
2/2004 Second International Green Processing Conference, Fremantle, WA, May 2004
3/2004 EXPLO 2004, Perth, WA, July 2004
4/2004 Metallurgical Plant Design and Operating Strategies, Perth, WA, September 2004
5/2004 PACRIM 2004 Congress, Adelaide, SA, September 2004
6/2004 Bac-Min Conference, Bendigo, Vic, November 2004
7/2004 Orebody Modelling and Strategic Mine Planning, Perth, WA, November 2004
8/2004 Coal2004, Fifth Australasian Coal Operators’ Conference, Wollongong, NSW,
February 2004
2005 1/2005 Ninth Underground Operators’ Conference, Perth, WA, March 2005
2/2005 Coal2005, Sixth Australasian Coal Operators’ Conference, Brisbane, Qld, April
2005
3/2005 New Leaders 2005, Brisbane, Qld, April 2005
4/2005 Second World Conference on Sampling and Blending, Sunshine Coast, Qld, May
2005
5/2005 Centenary of Flotation Symposium, Brisbane, Qld, June 2005
6/2005 Eighth International Mine Ventilation Congress, Brisbane, Qld, July 2005
7/2005 Hoist and Haul Conference, Perth, WA, September 2005
8/2005 Iron Ore Conference, Fremantle, WA, September 2005
9/2005 First Extractive Metallurgy Operators’ Conference, Brisbane, Qld, November 2005

Copies of all books currently in print can be obtained from The Institute office
Tel (03) 9662 3166 or Fax (03) 9662 3662

* = Out of print
The ‘S’ numbers in the third column refer to an older identifying number for Symposia, the numbers preceeding the ‘S’ number signify the new
publication ordering number.
2006 1/2006 Disposal of Mining Waste Symposium, Sydney, NSW, March 2006
2/2006 New Leaders 2006, Kalgoorlie, WA, April 2006
3/2006 Third International Green Processing Conference, Newcastle, NSW, June 2006
4/2006 Coal2006, Seventh Underground Coal Operators’ Conference, Wollongong,
NSW, July 2006
5/2006 Australia’s Uranium, Adelaide, SA, July 2006
6/2006 Sixth International Mining Geology Conference, Darwin, NT, August 2006
7/2006 AusIMM New Zealand Branch 2006 Conference, Waihi, New Zealand, August
2006
8/2006 Metallurgical Plant Design and Operating Strategies, Perth, WA, September 2006
9/2006 International Mine Management Conference, Melbourne, Vic, October 2006
10/2006 Water in Mining 2006, Brisbane, Qld, November 2006
2007 1/2007 Ninth Mill Operators’ Conference, Fremantle, WA, March 2007
2/2007 New Leaders 2007, Brisbane, Qld, May 2007
3/2007 Australia’s Uranium, Darwin, NT, May 2007
4/2007 Project Evaluation Conference, Melbourne, Vic, June 2007
5/2007 AusIMM New Zealand Branch 2007 Conference, Christchurch, New Zealand,
August 2007
6/2007 Iron Ore Conference, Perth, WA, August 2007
7/2007 EXPLO 2007, Wollongong, NSW, September 2007
8/2007 6th Large Open Pit Mining Conference, Perth, WA, September 2007
9/2007 World Gold 2007, Cairns, Qld, October 2007
10/2007 Uranium Reporting Workshop, Adelaide, SA, November 2007
2008 1/2008 Coal2008, Eighth Underground Coal Operators’ Conference, Wollongong, NSW,
February 2008
2/2008 Tenth Underground Operators’ Conference, Launceston, Tas, April 2008
3/2008 13th Australian Tunnelling Conference, Melbourne, Vic, May 2008
4/2008 Sampling 2008 Conference, Perth, WA, May 2008
5/2008 AusIMM International Uranium Conference, Adelaide, SA, June 2008
6/2008 The AusIMM New Leaders’ Conference 2008, Wollongong, NSW, July 2008
7/2008 MetPlant 2008, Perth, WA, August 2008
8/2008 Ninth International Congress for Applied Mineralogy (ICAM) 2008, Brisbane,
Qld, September 2008
9/2008 Narrow Vein Mining Conference 2008, Ballarat, Vic, October 2008
10/2008 First International Future Mining Conference and Exhibition 2008, Sydney,
NSW, November 2008
11/2008 PACRIM Congress 2008, Gold Coast, Qld, November 2008
2009 1/2009 Coal2009, Ninth Underground Coal Operators’ Conference, Wollongong, NSW,
February 2009
2/2009 Orebody Modelling and Strategic Mine Planning, Perth, WA, March 2009
3/2009 Project Evaluation 2009 Conference, Melbourne, Vic, April 2009
4/2009 The AusIMM New Leaders’ Conference 2009, Brisbane, Qld, April 2009
5/2009 AusIMM International Uranium Conference, Darwin, NT, June 2009
6/2009 Sustainable Development Indicators in the Minerals Industry (SDIMI) 2009,
Gold Coast, Qld, July 2009

Copies of all books currently in print can be obtained from The Institute office
Tel (03) 9662 3166 or Fax (03) 9662 3662

* = Out of print
The ‘S’ numbers in the third column refer to an older identifying number for Symposia, the numbers preceeding the ‘S’ number signify the new
publication ordering number.
 7/2009 Iron Ore 2009 Conference, Perth, WA, July 2009
8/2009 Seventh International Mining Geology Conference 2009, Perth, WA, August
2009
9/2009 AusIMM New Zealand Branch Conference 2009, Queenstown, New Zealand,
August 2009
10/2009 Water in Mining 2009, Perth, WA, September 2009
11/2009 Tenth Mill Operators’ Conference, Adelaide, SA, October 2009
2010 1/2010 Tenth Underground Coal Operators’ Conference, Wollongong, NSW, February
2010
2/2010 The AusIMM New Leaders’ Conference 2010, Kalgoorlie, WA, April 2010
3/2010 Sampling 2010 Conference, Perth, WA, May 2010
4/2010 AusIMM International Uranium Conference, Adelaide, SA, June 2010
5/2010 Seventh Large Open Pit Mining Conference, Perth, WA, July 2010

Copies of all books currently in print can be obtained from The Institute office
Tel (03) 9662 3166 or Fax (03) 9662 3662

* = Out of print
The ‘S’ numbers in the third column refer to an older identifying number for Symposia, the numbers preceeding the ‘S’ number signify the new
publication ordering number.
 SPECTRUM SERIES
1. Making the Mount Isa Mine, 1923 - 1933 Don Berkman 1996
2. History of Drilling Graham McGoggan 1996
3. The Cobar Mineral Field – A 1996 Perspective Warren Cook 1996
Andrew Ford
Julian McDermott
Peter Standish
Craig Stegman and
Therese Stegman
4. Towards 2000 – Resource to Reserve Inputs Seminar, Melbourne, Vic 1997

5. Towards 2000 – National Conference on Ironmaking Resources and Reserves 1997

Estimation, Perth, WA
6. Towards 2000 – The Resource Database Towards 2000, Wollongong, NSW 1997
7. Towards 2000 – Ore Reserves and Finance, Sydney, NSW 1998
8. Towards 2000 – Assessment of Reserves in Low Rank Coals, Morwell, Vic 1997
9. Towards 2000 – Ore Reserve Reconciliation Workshop, Darwin, NT 1997
10. Towards 2000 – Gold and Nickel Ore Reserve Estimation Practice Seminar 1998
11. Towards 2000 – Resource/Reserves Estimation Practice in the Central West 1999
New South Wales Mining Industry, Cobar, NSW
12. Field Guide for Geoscientists and Technicians
1st Edition 2004
2nd Edition 2007
13. The Extractive Metallurgy of Zinc Roderick J Sinclair 2005
14. Orebody Modelling and Strategic Mine Planning – Uncertainty and Risk Roussos Dimitrakopoulos
Management Models
1st Edition 2005
2nd Edition 2007
15. The Extractive Metallurgy of Lead Roderick J Sinclair 2009
16. Flotation Plant Optimisation – A Metallurgical Guide to Identifying and Ed: Christopher J Greet 2010
Solving Problems in Flotation Plants
17. Advances in Orebody Modelling and Strategic Mine Planning I – Old and Roussos Dimitrakopoulos 2010
New Dimensions in a Changing World

The Global Optimiser Works – What Next?
J Whittle1
ABSTRACT
The Global Optimiser used by Whittle Consulting has gone through three
major versions to date. The first was based on the Milawa optimisation
algorithm; it worked, but had many shortcomings. The second, known
internally as ProberA, had a different approach to optimisation in that it
used a series of random starting points and found the nearest local NPV
maximum to each. It did this a sufficient number of times to give us some
confidence that the best result found was close to optimal. ProberB was
an enhanced version of ProberA, with the ability to handle a wider range
of constraints, particularly with regard to limits on the differences in
depth between adjacent areas of a pit. ProberB has been used successfully
for some time now. It has produced excellent Life of Project schedules for
a wide range of very large mining complexes. However, like any piece of
software, it has its limitations. For example, it only copes directly with
three steps – mining, processing, and blending.
It is possible to ‘fool’ the program into handling other steps, but only
by using mental and mathematical gymnastics. This paper reviews the
mechanisms behind the Prober series and describes the plans for the next
version – ProberC.
INTRODUCTION
For some years now a global optimiser has been used to produce
long-term mining and processing schedules which maximise the
Net Present Value (NPV) of the whole project, taking account of
all cash flows including capital expenditure. There have been
three versions of this optimiser, known internally as Z3, ProberA
and ProberB. Further details can be found in Whittle and Whittle
(2007) and in Whittle (2007). This paper explains the methods
used in Prober and describes the plans for the development of a
new optimiser – ProberC.
MODELLING
In order to use a computer to optimise any system it is necessary
to create a mathematical model of the system. This model must
represent the system as accurately as possible, but also be
amenable to optimisation. In theory any mathematical model can
be optimised by throwing enough computer power at it.
However, as the size of models increases, some types can get
rapidly out of hand. There are mathematical problems which are
quite easy to describe but could not be fully optimised by using
all the computers in the world for a million years. However, it is
sometimes possible to find very good near-optimal solutions to
problems that defy full optimisation. Occasionally this can be
done by simplifying the model, other times by using approximate
optimisation methods. The Prober approach uses both these
techniques to obtain good, long-term schedules for large mining
complexes. Keep in mind that all models are wrong but some are
useful (Box, 1979).
A large mining complex with 20 or more pits with associated
processing plants and infrastructure would defy Life of Project
NPV optimisation if modelled in day-to-day detail. It is therefore
necessary to simplify the model to some degree. The main
simplification is to concentrate on long-term scheduling. If a
large expense is delayed by five years, this could have a
significant effect on the NPV. If the same expense is delayed by a
week, the effect will be small. Long-term scheduling decides
what will be mined, processed, blended and sold in each year of
1. FAusIMM, Director of Research, Whittle Consulting Pty Ltd,
42 Yeneda Street, Balwyn North Vic 3104.
Email: jeff@whittle.name
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
the life of the complex. Expenses are delayed and revenues are
brought forward, while meeting all the required constraints. In
short-term scheduling, practicality and convenience are more
important than the timing of cash flows. There is also the
consideration that the data required for short-term scheduling
differs from that for long-term scheduling. For long-term
scheduling what matters is the total volume of dirt a truck can
move in a year; however, there is no need to consider its detailed
movements.
For Prober, the resource is modelled as a number of
‘sequences’ of ‘panels’, and the rule is that mining of a panel
can’t start until the panel before it in the sequence has been
completely mined. These panels usually represent benches in a
push-back, but can also represent parts of an underground mine
that must be accessed in a particular order. A panel consists of
one or more ‘parcels’ of material, each of which will typically
have a material type, a tonnage and a number of grades. If Prober
mines only a fraction of a panel in a particular year, it uses the
mathematical fiction that the same fraction of every parcel in the
panel will be mined in that year. Of course, this is not what
will happen in the day-to-day scheduling, but the necessary
adjustments should have little effect on the NPV.
The main data input to Prober consists of a ‘resource tree’
showing all possible things that can happen to material after it
has been mined. For each parcel, it gives the alternative
processing paths that can be taken (eg mill or heap leach). For
each processing path, it gives the products that the outputs from
processing can be blended into, etc. The remainder of the data
details cost and price factors apply to the whole complex,
together with the various operational constraints. When
modelling the constraints on the operation of the complex, it is
not necessary to simplify much at all. Limits can be set on
mining and processing throughputs, which can depend not only
on tonnage but on grades or other characteristics (eg how much
power or acid is required to process material). Upper and lower
limits can be set for the average grades input to processing and in
products, as well as on the depth separation between sequences.
THE PROBLEM OF LOCAL MAXIMA
The aim is to find the mining and processing schedule which
gives the highest NPV. Consider just one sequence and one year
with no blending. This might result in the situation illustrated in
Figure 1. Sequence A has some panels, some of which are ore
and others are waste. As mining proceeds through the ore, the
value goes up; as it proceeds through waste, it goes down. The
result in this case is that there are two peaks. Figure 2 illustrates
the same idea for another sequence. Each peak is a maximum in
its own part of the graph, but it is a local maximum. Only one is
the global maximum. Figure 3 shows what happens if the
complexity is increased to two sequences, but still only over one
year. Now there are four local maxima. In practice, with a few
dozen pits, blending constraints and many years, there can be
thousands of local maxima. This can be equated to the search for
the highest point in the Himalayas rather than the highest point
of Mount Fuji. Merely going uphill from a random start point
will get you to a ‘nearest’ local maximum. It is unlikely to be the
highest point, that is, the global maximum. Because of the
multiplicity of local maxima, this is not a linear problem in the
sense of linear optimisation – it is a ‘mixed integer’ problem.
Spectrum Series Volume 17 3
J WHITTLE
FIG 1 - The total value of material mined from Sequence A plotted
against depth mined, where some benches are ore and some
are waste. The two peaks are local maxima; the left peak is the
global maximum.
FIG 2 - The total value of material mined from Sequence B plotted
against depth mined, where some benches are ore and some are
waste. The two peaks are local maxima; the right peak is the
global maximum.
FIG 3 - The total value of material mined when both Sequence A
and Sequence B are mined to various depths. There are four peaks.
Only the rear left peak is the global maximum. This shows just two
sequences and one year. In real cases with many sequences and
many years, there can be thousands of local maxima.
Although linear optimisation methods are very fast for large
problems, the introduction of integers – which is necessary to
deal with the multiple local maxima – slows the process down by
several orders of magnitude. Indeed, schedules for very large
mining complexes are effectively unobtainable using mixed
integer software.
4 Spectrum Series Volume 17
HOW THE PROBER SERIES OF PROGRAMS
WORK
The Prober series of programs break the problem down into a
number of smaller problems which can each be solved quickly
using linear optimisation software. Prober first produces a
random feasible mining schedule consisting of a year-end depth
in each sequence for each year in the life of the project. Thus, for
about 20 sequences and 15 years, there would be 300 depths.
This schedule is produced by using pseudo-random numbers to
calculate the depths while taking account of all the constraints on
mining throughput and on depth separation. This fixes the
material available for processing in each year. There is no
optimisation in this step and this is not necessarily a good
schedule. Indeed, it is usually very poor, but it is feasible from a
mining point of view, so it gives us a starting point. You can
think of this schedule as a random point on the ground
somewhere in the Himalayas. Given the material made available
in each year, it is a straightforward matter to use standard linear
optimisation to determine what to process, blend and sell so as to
maximise the NPV obtainable by processing the material subject
to the throughput constraints.
As explained earlier, the assumption is made, so far as the
model is concerned, that parcels within a panel are all mined in
the same proportion. This means that changes in mining depth
within a panel have linear effects on the quantities of material
available for processing and so on. Consequently, in addition to
having the linear optimisation package control what to process,
blend and sell, it is possible to let it control the depths mined,
providing that those depths are constrained to stay within the
panel that they started in. The result is the best mining and
processing schedule for a given set of year-end panels. Having
obtained such a result, the depth details suggest changes to the
year-end panels. For example, if the year-end panel for a
particular sequence and period is set to three, but the
optimisation does not mine any of Panel 3, the program will
change the year-end panel to two and try again. If the NPV
improves, the process is repeated. The program is effectively
going uphill in the NPV landscape, and this process is repeated
until a maximum is reached. This is the nearest local maximum
to the initial random feasible schedule. One of the peaks in the
Himalayas has been found.
The Prober series of programs works by repeatedly creating a
random feasible solution and then finding the nearest local
maximum. The various NPVs that the program finds are kept track
of, and the run is usually stopped when the top ten values lie
within 0.1 per cent of each other. This does not guarantee
optimality, but it does give confidence that the best results are
pretty close. Most projects that we worked on had a distribution of
local maxima NPVs that is pretty flat near the highest values as
illustrated in Figure 4, which is from a real case. That is, many
different schedules are found that have NPVs which are very
similar to the best. In this case, 11 of the 500 local maxima found
are in the top 0.1 per cent and over 200 are in the top one per cent.
This can be an advantage because it gives managers the
opportunity to choose the schedule that best suits less measurable
considerations such as social, political, environmental and project
risk effects. To date, many projects have been successfully
completed with the current version of Prober (ProberB) and very
significant NPV gains have been made totalling billions of dollars.
PROBERC
ProberC uses the same optimisation system but operates on a
greatly enhanced data structure. It can be regarded as a
generalisation of ProberB. Grades are expressed as quantities, so
that a grade of 1.2 grams/tonne of gold in a 2000 tonne portion of
material would be expressed as 2400 grams. Thus, a portion of
Advances in Orebody Modelling and Strategic Mine Planning I

FIG 4 - 500 NPVs obtained by moving from 500 random schedules
to the nearest local maximum have been sorted into ascending
order of value. The curve is relatively flat at the high end and 11 of
the local maxima have NPVs that are within 0.1 per cent of the best.
material, be it run-of-mine or the output from processing, is
described by the quantities associated with it. Any of these
quantities can be involved in throughput limits, costs, revenues,
etc. In ProberC, everything that is done to material is done by a
‘procedure’. Mining, processing, smelting, blending, etc are all
procedures. There can be as many procedures as required and
any procedure can provide input to another procedure, with the
restriction that material cannot ‘loop’ through the system and
thus provide input to the procedure that produced it. Thus, there
is an arbitrary tree structure of procedures, starting with a mining
procedure and going through as many steps as required to the
point(s) of sale or to waste. This contrasts with ProberB, which
handles only mining, processing and blending for sale. Delivery
of material from a procedure to stockpile, sale, discard or another
procedure is much more explicit in ProberC than ProberB and
has its own cost structure. For example, long distance delivery
costs can be handled explicitly rather than having to be included
in the processing cost as in ProberB.
All costs and revenues are calculated as the sum of one or
more of three components, multiplied by a scaling factor. The
components are:
1. a cost or revenue proportional to one or more of the
quantities, which describe the portion of material involved;
Advances in Orebody Modelling and Strategic Mine Planning I
THE GLOBAL OPTIMISER WORKS – WHAT NEXT?
2. a cost or revenue proportional to the difference between a
grade (ratio between two quantities) and a base grade
multiplied by a third quantity; and
3. a constant attached to each portion of material in the
resource tree each time it can be processed or delivered.
All the factors and base grades in components 1 and 2 can vary
independently with the year. Any number of components of types
1 and 2 can apply. Type 3 components are attached to each
portion of material operated on and can be calculated to suit that
portion, but are fixed in time. The scaling factor by which the
sum of these components is multiplied can vary with the year.
This system allows almost any cost or revenue structure to be
modelled. Setting up such costs and revenues might appear
somewhat daunting, but it is important to remember that the data
is always set up by a specialist consultant, not the end user. Costs
and throughput limits can be applied to both the inputs and the
outputs of procedures. Despite the great increase in complexity
of the processing and cost structures that can be modelled, the
problem steps can each still be handled by a standard linear
optimisation package as in ProberB. ProberC is currently under
development and it is expected to be completed in 2009.
CONCLUSIONS
ProberB has established itself as a powerful and flexible tool.
ProberC will have the same power as ProberB, but will handle a
much wider range of project and costing complexity.
REFERENCES
Box, G E P, 1979. Robustness in the strategy of scientific model building,
in Robustness in Statistics (eds: R L Launer and G N Wilkinson),
p 202 (New York: Academic Press).
Whittle, G, 2007. Global asset optimisation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 331-336 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Whittle, J and Whittle, G, 2007. Global long-term optimisation of very
large mining complexes in Proceedings APCOM 2007 – 33rd
International Symposium on Application of Computers and
Operations Research in the Mineral Industry, Santiago, pp 253-260.
Spectrum Series Volume 17 5

Optimal Mining Principles
B King1
ABSTRACT
Picture yourself responsible for the exploitation of a world class deposit,
creating staggering quantities of products for global consumption over
decades, employing thousands of people, about to spend billions of dollars
on infrastructure and you are going to design this project. Should it be
surface or underground (or both), how big should the processing plants be,
what technology should be utilised, what happens if more resources are
found or the price forecast changes? This paper aims to help guide
engineers faced with the prospect of determining optimal mining policies
for large projects. It draws on experiences at some of the largest mining
projects and mining companies in the world, including the Bingham
Canyon Mine (USA), Freeport (Irian Jaya), Escondida (Chile),
Chuquicamata (Chile), Hamersley Iron (Australia), Ekati Diamond Mine
(Canada) and several Hunter Valley coal mines (Australia). The paper
outlines some important areas to put in place before starting, principles to
guide the planning process and suggestions for finding additional value.
INTRODUCTION
Terminology
One of the first sources of confusion in strategic planning arises
due to the definition of terms (or lack thereof). The terms used in
this paper are based on Ken Lane’s cut-off grade optimisation
work (Lane, 1964, 1988), which was based on Richard Bellman
foundational dynamic programming methodology (Bellman,
1957). Over the last ten years of working with major mining
projects and feasibility studies, I have found these definitions
useful and applicable in mining and other industries. Mining
projects have many choices and therefore many decisions to
make. A specific decision that is made in every period is grouped
and called a policy. For example, an open pit copper mine may
have a decision on mill cut-off grade to be made in every period.
When the same decision is made at every time period, the result
is a constant policy. Typically, different decisions are made over
time, in which case a variable policy is formed. Other operating
policies that may change over time would normally include how
much material is moved from a certain area, SAG mill grind size
and flotation residence time; each needs to be made in every
period.
Mining studies must typically determine many different
policies during the project life. A group of policies is referred to
as a strategy. For example, mining policies for a number of
pushbacks could be described as a mining strategy. By
combining a number of mining and processing policies, we can
arrive at the strategy for the project, or a strategic plan. A
decision, policy or strategy does not imply any optimisation
criteria have been used to define them – it simply defines a
number of decisions. Sadly, in real projects, many ‘strategic’
plans also have very little optimisation applied. In order to
‘optimise’ a decision (or policy or strategy), we need to have an
objective or objective function. The objective will be discussed
later in this paper, but for now, let’s assume it is some definition
of shareholder value. An optimisation process is a way of making
decisions to achieve the project objective. For example, a cut-off
grade optimisation process will optimise the cut-off grade policy
so that the Net Present Value (objective function) is maximised.
1. Managing Director, Strategy Optimisation Systems Pty Ltd, 24 Harris
Street, Wellington Point Qld 4160.
Email: Brett.King@Stops.com.au
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
There is a subtle distinction between optimising and
maximising. Decisions, policies and strategies are optimised (not
maximised), Net Present Value (NPV), cash flow or value is
maximised (not optimised). If one were to maximise the cut-off
grade (to an extreme), then no ore would be sent to the mill and
the maximum NPV is unlikely to be achieved. On occasion,
maximising a policy may also maximise the objective – it is
better to use the right terminology and sound like an expert!
The focus of this paper will be to utilise optimised strategies to
create optimised strategic plans. Given that all strategies discussed
in this paper will refer to optimal mining plans, a Strategic Plan in
this paper is defined as the process of making a number of
operating decisions over the project life, to achieve the project
objective. There is clearly a need in the industry to standardise our
terminology and develop an optimal mining practice for the
community. Ideally, this will be a collaborative effort between
mining companies and the service companies that support them.
These groups need to span the industry, to help reduce the
misunderstanding that often leads to poor exploitation of the
earth’s finite resources (King, 2008). The paper demonstrates these
principles using open pit metalliferous projects, although these are
actually generic principles that are appropriately applied to coal,
iron ore, diamonds, bauxite, gold and also many non-mining
projects.
Objectives
The first issue that needs to be clarified at the start of a planning
exercise is the work objective. Although there would appear to be
a clear and simple answer to the objective, the reality is often
blurred by conflicting instructions. Maximising shareholder
value is normally the objective of large and small mining
companies (and non-mining companies). This objective is
clarified by the major mining houses as maximising the NPV,
subject to a number of constraints such as safety and good
stewardship of the environment:
• NPV is a sound basis for evaluating companies outside of the
mining world. For example, stock market analysts will often
value the share price for a company (in any industry) by
calculating the NPV of all of the company assets, then
dividing it by the number of shares issued (Brealey and Myers,
2000). NPV includes a balanced valuation of short- term value
(which receives very little discounting) and long- term value
(which receives greater discounting), making it an appropriate
instrument for the commercial valuation of most businesses.
The main difficulties reported with valuing a project based on
NPV include uncertain information and Net Present Value
versus present value.
• With regards to uncertain information, typically there is
substantial uncertainty in key drivers of the NPV including
prices, costs, reserves and productivities. For example, price
is normally a huge driver of value, and financial analysts
recognise that price is related to supply and demand.
Theoretically, as different strategies are evaluated with
different quantities of metal produced, the prices could
change for each case. The complexity does not stop here –
the price is dependent on the supply from competitors, which
would require an analysis of what they are doing as well. If
this is not complex enough, the price is also dependent on the
demand for end products. This means that a comprehensive
estimate of price also requires analysis of what competitors
Spectrum Series Volume 17 7
B KING
and customers are planning. A risk/reward analysis would
normally indicate that a comprehensive competitor/customer
analysis is not valuable. Although specific aspects of risk are
often incorporated using sensitivity studies of targeted issues,
there is a still substantial development required to turn this
analysis into useful decision making tools for the mining
industry. Option value, and the ability to use variability to
add value and control risk, is regularly missing from project
valuation analysis. Advances in mathematical formulation
and simulation methodologies can reveal a more complete
understanding of project value. Incorporation of uncertainty
and risk needs to become common place in the industry for
fair valuation between investment alternatives.
The effort required to truly calculate a ‘Net’ Present Value
means that often important components are ignored and a present
value is used instead. Key aspects such as period costs, expansion
capital, sustaining capital and closure costs are not able to be
incorporated in some optimisation algorithms. Ignoring such
issues or modelling them inadequately means that the results are
crude, and value destroying decisions can be made.
While both of these difficulties appear to limit the suitability
of using NPV to value a project or different scenarios within a
project, NPV still remains at the heart of project valuations and is
normally superior to the other financial instruments such as
internal rate of return, payback period and simple cash flows.
Care should be taken when using other measurements to
approximate the value of a project. People not aware of NPVs
balancing of short- and long-term cash flows may be presented
with some of the following alternatives for the planning
objectives:
• product quantity – especially in the budgeting period (eg
gold or copper produced in the first three to five years,
• mineral resources or reserves (especially stated to the stock
market),
• smooth production schedules (eg constant material movement
and constant grades for processing),
• minimising costs or maximising short-term cash flow, and
• maximising employment.
Using the above alternatives for the objective will normally
result in a different plan than the one created by maximising
NPV and so will have a lower NPV. Take the objective of
maximising cash flow as an example. Waste stripping incurs
substantial costs that reduce the cash flow. If waste stripping is
suspended, cash flow for the first few years may be improved.
Eventually, there will be no ore available to process, the mill
capacity will not be utilised, and the cash flows (and NPV)
rapidly fall. The NPV can be used to balance the early costs of
waste stripping with the benefit of having the mill capacity
continuously utilised. All aspects of mining and downstream
processing that involve costs or revenue are captured directly in
the objective function (NPV). The better our estimates of these
costs and revenues, the better our estimates of NPV, which will
in turn help bring value to shareholders.
Constraints
If an aspect of the project does not have a cost or revenue it will
not directly be added to the project NPV, hence the inevitable
constraints that must be considered. There are many issues to be
considered in strategic planning that are not captured directly in
the objective function. Safety and environmental care are two
priorities of modern responsible mining that have important issues
that are not always evident in cost or revenue. These and many
more issues are captured in the optimisation process as constraints
to bound the schedules that are considered valid, and are known as
the feasible solutions. Many processes have a combination of
8 Spectrum Series Volume 17
costs, revenue and constraints. The safe operation of a mine or
processing plant will involve additional cost components and
constraints on how the activities can be undertaken. Typically
the constraints will involve restrictions in productivity rates.
Good environmental stewardship may involve continual ground
rehabilitation and water treatment costs – it may also require some
constraints such as tailings dam or waste dump footprints. Some
care is needed to make sure a cost is not mistakenly implemented
as a constraint.
Creating ‘unconstrained’ schedules may be of some temporary
benefit to understand the value drivers for a project, but be wary
of releasing such plans to a wider audience as it may generate
unrealistic expectations. For example, when considering the size
of a process plant expansion, the mining rates may initially be
left unconstrained. This unconstrained case may be used to get
the maximum size of the truck and shovel fleets. However, the
best process plant capacity, when unconstrained, may not be the
best process plant capacity when constrained. The unconstrained
cases would tend to over-expand the mill capacity since it does
not consider the purchase costs for the mining equipment. These
‘unconstrained’ schedules and their ‘imaginary’ values are not
achievable and so would only be used for finding the boundaries
for analysis.
Plan the planning
There could well be only one chance to determine the best value
of a project, then acquisition, disposal or operation. Time, staff,
consultants, computing power and budgets are all limited
resources that need to be used wisely. Much like the optimisation
process we apply to a project, we need to identify the resources
needed for some analysis and make sure that key people and
support is available during the study. Experience suggests that
very few people have been exposed to this field of the business,
and so the need for training before the study is necessary, as well
as support for when things do not simply fall into place. One of
the problems with using inexperienced people is that they often do
not know when additional value is possible from a project. The
new planner can make assumptions that are not easily seen in the
resulting schedules and so may not be picked up by management
or peer review.
The unrealised value ‘left on the table’ by inexperienced
analysis can easily be several percentage points of the NPV, and
for large multi-billion dollar projects these are substantial values.
It is therefore essential to have the best people working on these
projects and have them reviewed by strategic planning specialists
(internal and/or consultants). Real world strategic planning is
normally constrained by time. The time spent designing a model,
collecting and validating inputs, building the model, running
cases, analysing the cases and then presenting the results, these
activities consume precious and finite time resources. Experience
would suggest that gathering and validating inputs often increases
as input data is collected and checked. Using available information
such as budget costs might initially appear to rapidly provide
optimisation inputs, but they may need to be substantially re-
categorised into fixed and variable components for strategic
planning purposes. A planned work program will have a
contingency for some unforseen data collection issues and changes
in direction due to initial analysis. Typically, a Strategic Planning
Work Program would have the components in Table 1.
For new greenfield or acquisition projects, approximately half
the available time is typically used to collect appropriate data,
validate it and get it into the appropriate format for use by the
planning software. Once a model is validated, a new expansion
study, or the incorporation of additional reserves, requires much
less data validation and the emphasis shifts to the analysis of
results. Several of the tasks in Table 1 are done in parallel due to
the iterative nature of the planning process. Note that Table 1
does not include the training time to ensure people understand
what to do in each of the phases.
Advances in Orebody Modelling and Strategic Mine Planning I

TABLE 1
Summary of typical strategic planning effort.
Tasks New model Existing model
Scoping and understanding the 10% 10%
existing project
Gathering, validating and 50% 20%
transforming data
Model building 10% 10%
Running scenarios and 20% 50%
analysing the results
Reporting on results 10% 10%
Given how long data validation and analysis takes, it is critical
to plan a realistic scope. ‘Dry run’ strategic planning exercises
can help identify training requirements, holes in data
requirements, streamline data acquisition and align scope
expectations. There are typically too many options to evaluate
and not enough time, so the planned activity should focus on
identifying the highest value controllable drivers. For example,
although blasthole initiation systems may be expensive and have
room for reduction, they may not be as important to consider as
shovel productivity as a driver of project value. Some value
drivers may not be controllable, such as taxation, statutory
compliance fees, award wages, cost of power and possibly even
the commodity price. Although these uncontrollable drivers are
fixed in the analysis, they should be considered in the NPV
calculations, since they may influence decisions.
Simplifying assumptions
Large mineral resources may take many years or even decades to
fully exploit and have more decisions to make than there are
electrons in the universe (approximately 1079). The currently
available algorithms for mine design, scheduling and processing
require simplifying assumptions to find a solution and substantial
computational power to do any sort of partial optimisation. For
example, the Lerchs-Grossman (LG) algorithm, used in finding
ultimate pit limits (Lerchs and Grossmann, 1965) – fundamentally,
this algorithm uses the assumption that we know the value of
every block in a model. This is somewhat difficult when we do not
know when the block is to be mined, what the price will be at that
time, what other blocks are being mined with this one, what
processes are available or whether the ideal process will have
capacity for the block. This is a simplification in order to estimate
the shape of the mine when completely mined.
If the value is dependent on time or capacity-related issues, then
this casts doubts as to just how ‘optimal’ the LG results are. This is
not to say that the algorithms cannot be used to guide the best
location of the ultimate pit limits, it must however be taken into
account when interpreting the results. Phrases with ‘optimise’ and
‘plans’ should be regarded with a healthy scepticism – we should
also realise that we can rarely claim to have the optimum plan,
rather something closer to the highest value plan we could deliver
in the time allowed. The algorithms underlying all currently
available software tools have limitations due to their assumptions
– skilled engineers are able to use them appropriately in spite of
these limitations. Some other simplifications typically made in
strategic planning include:
• mineral resource estimates accurately reflecting what mining
and processing will deliver;
• annual equipment productivity, utilisation and availability
estimates;
• predictable costs and prices over the life of the project; and
• pretax and post-tax optimisation leading to the same
decisions.
Advances in Orebody Modelling and Strategic Mine Planning I
OPTIMAL MINING PRINCIPLES
For annual plans that cover the entire life of a major mineral
resource, we will not typically model individual salaries, weekly
maintenance schedules, shovel loading configurations or the
extra day that occurs in leap years. We must draw the distinction
between different planning time frames (real time, shift, weekly,
monthly, quarterly and annual) and the appropriate level of detail
for the analysis.
As a general procedure, it is advisable to start off with a simple
model of the entire project, then add complexity if required. Once
you have a simple model it can be used and its results analysed.
When a component of the cash flow calculation is found to be
poorly approximating the real value, it can be modelled in
greater detail. For example, an initial model of a mine may start
with the simple mining cost per mass of material moved. This
model then may be broken down into components and their
drivers, such as drilling ($/t), blasting ($/t), loading ($/m3) and
hauling ($/truck_operating_hour). The haulage component may
require further detail by modelling the source elevation,
destination of the material and the height of the dumps and so on.
The objective is to arrive at a good estimate of the NPV for
several different alternatives. If a simple model provides a good
estimate, then there is no need to pursue a complex model.
Another advantage of this approach is that some analysis can be
done quickly into the project – this may enable the project scope
to be tailored to the value drivers as they are better quantified. It
is often surprising how simple models can estimate value so
accurately and drive the decision process appropriately.
Knowledge and understanding – keys for
discovering value
This section of the paper aims to point people in directions to find
more value for their projects. It appears that no one person or
company has exclusive access to value creating ideas. It is my
conviction that many high value ideas will flow through people
who intimately understand a project, not just through the industry
expert who comes in for a three day review. A sound knowledge of
the project, value drivers and optimisation assumptions are all
important to unlocking greater value from a mining project. The
variety of geological resources, mining methods and philosophies
employed at major projects around the globe mean that value
discovered in one project may not be found in another. The
following suggestions are general areas that are regularly
neglected and so may be sources of additional value.
Project optimisation
Of all the optimisation analysis that is undertaken in geology,
mining, processing and marketing, it seems that very little analysis
spans all of these disciplines. Although the same detail cannot be
applied for global project analysis as can be applied to detailed
component analysis, we need to start with the big picture to make
most effective use of our finite resources (such as time, people and
budgets). The project analysis allows for determination of where
the most important value drivers are, allowing prioritising of
analysis effort. The project optimisation can also provide context
and scope for a more detailed analysis.
Working out of your discipline
Concentrated and isolated analysis within geology, mining,
processing, environmental, marketing and financial can easily
develop silos of knowledge and understanding. While this may
start at university and a natural affinity with one’s discipline, it is
essential for project optimisation that some people come to
understand the whole project. Many fields are interrelated – for
example, problems in the processing plant could well stem from
fundamentally different geological ore genesis. The cut-off grade
of material mined has substantial implications on processing and
marketing parameters.
Spectrum Series Volume 17 9
B KING
Multidisciplinary teams
Major mining projects are well beyond the technical skills of any
one individual. Creating teams to look for value has often led to
unrealised value. The quantity of information to be absorbed and
translated often works well with a multidisciplinary team. Teams
would normally include geology, mining, processing, environ-
mental and financial members and it is essential that some of the
team members should also have a strong optimisation capability.
Well-run teams can encourage innovation and building on each
other’s ideas – they are also often able to quickly discard options
unlikely to add value.
Removing constraints
Because constraints place substantial limits on the value of a
project, what is the value of removing these constraints? The
concept of de-bottlenecking or using constraints theory may be
used, as long as it is in conjunction with value analysis and not
just for gaining additional capacity. If the primary crusher is a
major bottleneck, would the value obtained by adding capacity
cover the new capacity?
Constraints or targets
The direction of a project can easily take on the preconceived
ideas of an influential manager. Creating smooth ore-grade
profiles has a cost to the mine – what is the value implication if
the processing plant accepted more variability? Marketing
‘constraints’ may be little more than market predictions; are
these creating ceilings on increasing the value of a project? Many
of these issues are quickly identified by multidisciplinary teams.
Full value focus
Focus on the corporate objective of what adds value, even if that
does not suit some parties. Utilise the whole of the resource
rather than just the first 20 to 30 years (it will help demonstrate
the value of expansions). Making an open pit bigger should be
considered in the light of the alternate underground options. How
can the project realise the greatest value from one of these
options? Don’t just use average quantities – know the value of:
• a truck of ore,
• a hole of explosives,
• a conveyor belt,
• SAG steel balls, and
• acid and wheel motors.
The more people understand the value, the more they are likely
to make appropriate value based decisions. Break-even cut-off
grade costs calculations for reclaiming stockpiles are presented
in an appendix of this paper to illustrate the need and value of
understanding optimisation assumptions.
CONCLUSIONS
This paper has provided guidance to help strategic planning
engineers to fully and profitably exploit the resources in their
care. Shareholder value within a constraint framework is
normally the objective for optimal mine plans – the key is to not
get so distracted by details that focus is lost on what really adds
value. These are finite resources that we normally only get one
chance to exploit – let’s make the most of them while achieving
the best value for shareholders.
REFERENCES
Bellman, R, 1957. Dynamic Programming (Princeton University Press:
New Jersey).
Brealey, R A and Myers, S C, 2000. Principles of Corporate Finance,
sixth edition (Irwin McGraw-Hill).
10 Spectrum Series Volume 17
King, B, 1999. Cash flow grades – Scheduling rocks with different
throughput characteristics, in Proceedings Whittle Strategic Mine
Planning Conference, pp 103-109 (Whittle Programming: Melbourne).
King, B, 2008. Optimal mining – Building the practice of optimisation in
the mining industry [online]. Available from: <http://www.optimal
mining.org/optimalmining/>.
Lane, K F, 1964. Choosing the optimum cut-off grade, Q Col Sch Mines,
59(4):811-829.
Lane, K F, 1988. The Economic Definition of Ore – Cut-off Grades in
Theory and Practice (Mining Journal Books: London).
Lerchs, H and Grossman, I F, 1965. Optimum design of open-pit mines,
Transactions of the Canadian Institute of Mining, LXVIII:17-24.
Wooller, R, 2007. Optimising multiple operating policies for exploiting
complex resources – An overview of the COMET scheduler,
in Orebody Modelling and Strategic Mine Planning, second edition
(ed: R Dimitrakopoulos), pp 309-316 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
APPENDIX – APPLYING BREAK-EVEN
CUT-OFF GRADES
The following example has been presented to show how care must
be exercised when determining what costs should be included in
calculating operational cut-off grades. This example highlights a
number of scenarios involving reclaim from stockpiles and which
costs should be included in each case. Consider the costs that are
used to calculate a mill cut-off grade – should the mining (drill,
blast, load and haul) and G&A (General & Administration, $/year)
type costs be included for material going to or coming from a
stockpile?
First, take the example of a project that is ‘mine constrained’ –
the project is struggling to use the processing capacity due to
mining capacity restrictions (truck or shovel capacity, sinking
rates, etc). This will typically occur at the beginning of a project
when the high value material is only located at depth. It also may
occur after mining one high grade area before the next high grade
area is available. Even when using all the mining equipment, there
may not be sufficient high grade ore to fill the mill capacity.
What is the minimum grade that can be sent for processing?
The question can easily be answered by considering the two
options possible for a region of ground:
1. What is the value of sending the material to the mill?
2. What is the value of sending the material to waste (Waste
value will be a negative number since all components are
negative value costs)?
Since the material is being mined, when the value of sending
the material to the mill is greater than sending it to the waste
dumps, it should be processed. Using variable definitions as used
by Lane (1988) where possible, such as the period cost (f, $/year),
processing cost (h, $/t ore), mining cost (m, $/t rock), selling cost
(k, $/t metal), price (p, $/t metal), recovery (y, recovered metal/in
situ metal), mining capacity (M, t rock/year), mass above cut-off
grade (x, t ore/t rock) and grade (g, in situ metal/t rock), the
following relationships are clear:
VORE = (p-k).y.g.x – f/M – h.x - mORE
VWASTE = -f/M - mWASTE
The mining cost of processing material as ore (mORE) could be
different to that when the same material is processed as waste
(mWASTE). For example, some operations will blast the material
more coarsely if it is being sent to waste, and the haulage costs
are often different between ore and waste. The mine constrained
cut-off grade for material mined from the pit (gMP) is determined
when VORE = VWASTE as follows:
h + mORE − mWASTE
gM =
( p − k ). y
Advances in Orebody Modelling and Strategic Mine Planning I

Apparent from the derivations and the formula for gM above,
the period costs do not have to be covered in this situation. This
observation makes sense when one considers that the period
costs will be incurred irrespective of the destination of the
material. When the ore and waste mining costs are identical, they
also drop out of the calculation and it becomes the same equation
that Ken Lane formulated (1988, page 28). This shows that
material mined in a pit will incur mining costs regardless of its
final destination. The decision can simply be made depending on
whether the project makes more money from processing the rock
than sending it to the waste dump. Where a stockpile exists
during the same mine constrained period, the following equation
to calculate the mine constrained stockpile break-even cut-off
grade (gMS) simplifies to:
h + mORE
gMS =
( p − k ). y
The mORE cost for a stockpile normally only includes the
reclaim loading and haulage costs, since drilling and blasting costs
are not necessary. Notice that here the waste mining cost does not
incur any costs since the material can simply be left on the
stockpile. The stockpile re-handling costs must be covered by the
processing revenue in this case. This means that higher grade
material would need to be reclaimed from the stockpile than what
could be sent directly from fresh material mined in the pit. Ore
ranking grades need to be updated (such as the cash flow grades,
King, 1999) to take this into account as material is moved from the
pit into a stockpile.
A subtle assumption in the above cut-off grade derivations is
that the material will be processed. Let’s consider a stockpile at
the end of the project life. If the material is not processed, it is
left on the ground and the project abandoned (ignoring closure
costs and reclamation for the moment). It is also assumed that
the environment is no longer mine constrained, but rather we are
processing constrained (H, t ore/year), and all the material mined
from the stockpile is ore (x = 1). In this case we consider the
following equations:
VORE = (p-k).y.g – f/M – h - mORE
Advances in Orebody Modelling and Strategic Mine Planning I
OPTIMAL MINING PRINCIPLES
VWASTE =0
The stockpile cut-off grade (gs) at the end of the project life is
again determined when VORE = VWASTE as follows:
f
h + mORE +
gS = H
( p − k ). y
This is an interesting result because now the break-even cut-off
grade for processing the last material from the mine stockpile
must include the mining costs and also the period costs. If
material is added to a stockpile using mine constrained cut-off
grades, then there could be substantial uneconomic material
mixed in with the economic material and it is likely impossible
to separate. The situation has more issues to be considered when
the material mined is not at the end of the project life and an
opportunity cost exists. There is an opportunity cost to reflect
which material is required to fill the limited processing capacity,
and the timing to access the remaining material. For a processing
constrained stockpile, a formulation of the opportunity cost
defines an economic cut-off grade (gE) as:
f +F
h + mORE +
gE = H
( p − k ). y
In practice, this formulation is very simplistic since it needs to
take into account multiple sequences, several processes,
interaction between several ore processing constraints, revenue
from multiple elements and time varying issues. It does show
that there is not one simply universal answer for the inclusion of
period and mining costs in determining what material should be
mined from, or deposited to, a stockpile. By taking the time to
understand the project situation and algorithm assumptions,
appropriate material can be sent to stockpiles and reclaimed.
Although opportunity costs or remaining value estimates are
implemented in modern cut-off grade optimisation software tools
like Comet (Wooller, 2007), they may not find the additional
value without a model constructed to look for this value.
Spectrum Series Volume 17 11

Stochastic Mine Planning – Methods, Examples and Value in an
Uncertain World
R Dimitrakopoulos1
ABSTRACT
Conventional approaches to estimating reserves, optimising mine
planning and production forecasting result in single, often biased,
forecasts. This is largely due to the non-linear propagation of errors in
understanding orebodies throughout the chain of mining. A new mine
planning paradigm is considered herein, integrating two elements:
stochastic simulation and stochastic optimisation. These elements provide
an extended mathematical framework that allows modelling and direct
integration of orebody uncertainty to mine design, production planning,
and valuation of mining projects and operations. This stochastic
framework increases the value of production schedules by 25 per cent.
Case studies also show that stochastic optimal pit limits:
• can be about 15 per cent larger in terms of total tonnage when
compared to the conventional optimal pit limits, while
• adding about 10 per cent of net present value (NPV) to that reported
above for stochastic production scheduling within the conventionally
optimal pit limits.
Results suggest a potential new contribution to the sustainable
utilisation of natural resources.
INTRODUCTION
Optimisation is a key aspect of mine design and production
scheduling for both open pit and underground mines. It deals with
the forecasting, maximisation, and management of cash flows
from a mining operation and is the key to the financial aspects of
mining ventures. A starting point for optimisation in the above
context is the representation of a mineral deposit in
three-dimensional space through an orebody model and the
mining blocks representing it; this is used to optimise designs and
production schedules (eg Whittle, 1999). Geostatistical estimation
methods have long been used to model the spatial distribution of
grades and other attributes of interest within the mining blocks
representing a deposit (David, 1988). The main drawback of
estimation techniques, be they geostatistical or not, is that they are
unable to reproduce the in situ variability of the deposit grades, as
inferred from the available data. Ignoring such a consequential
source of risk and uncertainty may lead to unrealistic production
expectations (eg Dimitrakopoulos, Farrelly and Godoy, 2002).
Figure 1 shows an example of unrealistic expectations in a
relatively small gold deposit. In this example (Dimitrakopoulos,
Farrelly and Godoy, 2002), the smoothing effect of estimation
methods generates unrealistic expectations of net present value in
the mine’s design, along with ore production performance, pit
limits, and so on. The figure shows that if the conventionally
constructed open pit design is tested against equally probable
simulated scenarios of the orebody, its performance will probably
not meet expectations. The conventionally expected NPV of the
mine has a 2 - 4 per cent chance to materialise, while it is expected
to be about 25 per cent less than forecasted. Note that in a
different example, the opposite could be the case.
For over a decade now, a traditional framework has been used
when dealing with uncertainty in the spatial distribution of
attributes of a mineral deposit, as well as its implications to
downstream studies, planning, valuation, and decision-making.
Now, a different framework than the traditional has been
suggested and is outlined in Figure 2. Instead of a single orebody
1. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montreal QC H3A 2A7, Canada.
Email: roussos.dimitrakopoulos@mcgill.ca
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
FIG 1 - Optimisation of mine design in an open pit gold mine, net
present value versus ‘pit shells’ and risk profile of the
conventionally optimal design.
model as an input to optimisation for mine design and a ‘correct’
assessment of individual key project indicators, a set of models
of the deposit can be used. These models are conditional to the
same available data and their statistical characteristics, and all are
constrained to reproduce all available information and represent
equally probable models of the actual spatial distribution of grades
(Journel, 1994). The availability of multiple equally probable
models of a deposit enables mine planners to assess the sensitivity
of pit design and long-term production scheduling to geological
uncertainty (eg Kent, Peattie and Chamberlain, 2007; Godoy,
2010, in this volume) and, more importantly, empower mine
planners to produce mine designs and production schedules with
substantially higher NPV assessments through stochastic
optimisation. Figure 3 shows an example from a major gold mine
presented in Godoy and Dimitrakopoulos (2004), where a
stochastic approach leads to a marked improvement of 28 per
cent in NPV over the life of the mine, compared to the standard
best practices employed at the mine; note that the pit limits used
are the same in both cases and are conventionally derived
through commercial optimisers (Whittle, 1999). The same study
also shows that the stochastic approach leads to substantially
lower potential deviation from production targets, that is, reduced
risk. A key contributor to substantial differences is that the
stochastic or risk-integrating approach can distinguish between
the ‘upside potential’ of the metal content, and thus economic
value of a mining block, from its ‘downside risk’, and then treat
them accordingly, as further discussed herein.
Figure 2 represents an extended mine planning framework that
is stochastic (that is, integrates uncertainty) and encompasses the
spatial stochastic model of geostatistics with that of stochastic
optimisation for mine design and production scheduling. Simply
put, in a stochastic mathematical programming model developed
for mine optimisation, the related coefficients are correlated
random variables that represent the economic value of each block
being mined in a deposit, which are in turn generated from
considering different realisations of metal content. Note that the
second key element of the risk-integrating approaches is
stochastic simulation; the reader is referred to Mustapha and
Dimitrakopoulos (2010, in this volume) for the description of
a new general method for high-order simulation of complex
Spectrum Series Volume 17 13
R DIMITRAKOPOULOS
FIG 2 - Traditional (deterministic or single model) view and practice versus risk-integrating (or stochastic) approach to mine modelling, from
reserves to production planning and life-of-mine scheduling, and assessment of key project indicators.
FIG 3 - The stochastic life-of-mine schedule in this large gold mine
has a 28 per cent higher value than the best conventional
(deterministic) one. All schedules are feasible.
geological phenomena. The further integration of market
uncertainties in terms of commodity prices and exchange rates is
discussed elsewhere (Abdel Sabour and Dimitrakopoulos, 2010,
in this volume; Meagher, Abdel Sabour and Dimitrakopoulos,
2010, in this volume).
The key idea in production scheduling that accounts for grade
uncertainty is relatively simple. A conventional optimiser (any
one of them) is deterministic by construction and evaluates a
cluster of blocks, such as that in Figure 4a, so as to decide when
to stop mining, which blocks to extract when, and so on,
assuming that the economic values of the mining blocks
considered (inputs to the optimiser) are the actual/real values. A
stochastic optimiser, also by construction, evaluates a cluster of
blocks, but as in Figure 4b, by simultaneously using all possible
combinations of economic values of the mining blocks in the
cluster being considered. As a result, substantially more local
information on joint local uncertainty is utilised, leading to much
more robust schedules that also can maximise the upside
potential of the deposit (eg higher NPV and metal production)
and at the same time minimise downsides (eg not meeting
production targets and related losses).
To elaborate on the above, the next sections examine a key
element in the risk-integrating framework shown in Figure 2, that
of stochastic optimisation. The latter optimisation is presented in
two approaches, one based on the technique of simulated
annealing, and a second based on stochastic integer programming.
Examples follow that demonstrate the practical aspects of
stochastic mine modelling, including the monetary benefits.
14 Spectrum Series Volume 17
STOCHASTIC OPTIMISATION IN MINE DESIGN
AND PRODUCTION SCHEDULING
Mine design and production scheduling for open pit mines is an
intricate, complex, and difficult problem to address due to its
large-scale and uncertainty in the key parameters involved. The
objective of the related optimisation process is to maximise the
total net present value of the mine plan. One of the most
significant parameters affecting the optimisation is the
uncertainty in the mineralised materials (resources) available in
the ground, which constitutes an uncertain supply for mine
production scheduling. A set of simulated orebodies provides a
quantified description of the uncertain supply. Two stochastic
optimisation methods are summarised in this section. The first is
based on simulated annealing (Godoy and Dimitrakopoulos,
2004; Leite and Dimitrakopoulos, 2007; Albor Consequega and
Dimitrakopoulos, 2009); and the second on stochastic integer
programming (Ramazan and Dimitrakopoulos, 2007, 2008;
Menabde et al, 2007; Leite and Dimitrakopoulos, 2010, in this
volume).
Production scheduling with simulated annealing
Simulated annealing is a heuristic optimisation method that
integrates the iterative improvement philosophy of the so-called
Metropolis algorithm with an adaptive ‘divide and conquer’
strategy for problem solving (Geman and Geman, 1984). When
several mine production schedules are under study, there is
always a set of blocks that are assigned to the same production
period throughout all production schedules; these are referred to
as the certain or 100 per cent probability blocks. To handle the
uncertainty in the blocks that do not have 100 per cent
probability, simulated annealing swaps these blocks between
candidate production periods so as to minimise the average
deviation from the production targets for N mining periods, and
for a series of S simulated orebody models, that is:
N
⎛ S S
⎞
MinO = ∑ ⎜ ∑ θ *n (s) − θ n (s) + ∑ ω *n (s) − ω n (s)⎟ , (1)
n =1 ⎝ s=1 s=1 ⎠
where θ *n ( s ) and ω *n ( s ) are the ore and waste production targets,
respectively, θ n ( s ) and ω n ( s ) represent the actual ore and waste
production of the perturbed mining sequence. Each swap of a
block is referred to as a perturbation.
Advances in Orebody Modelling and Strategic Mine Planning I
 STOCHASTIC MINE PLANNING – METHODS, EXAMPLES AND VALUE IN AN UNCERTAIN WORLD

FIG 4 - Production scheduling optimisation with conventional versus stochastic optimisers: (a) single representation of a cluster of mining
blocks in a mineral deposit as considered for scheduling by a conventional optimiser; and (b) a set of models of the same cluster of blocks
with multiple possible values considered simultaneously for scheduling by a stochastic optimiser.

The probability of acceptance or rejection of a perturbation is
given by:
⎧1, if Onew ≤ Oold
⎪
Prob{accept} = ⎨ Oold − Onew . stochastically simulated orebody models. The proposed SIP model
⎪⎩e T
, otherwise
This implies all favourable perturbations (Onew ≤ Οold) are
accepted with probability 1 and unfavourable perturbations are
accepted based on an exponential probability distribution, where
T represents the annealing temperature.
The steps of this approach, as depicted in Figure 5 are as
follows:
1. define ore and waste mining rates;
2. define a set of nested pits as per the Whittle implementation
(Whittle, 1999) of the Lerchs-Grossmann (1965) algorithm,
or any pit parameterisation;
3. use a commercial scheduler to schedule a number of
simulated realisations of the orebody given 1 and 2;
4. employ simulated annealing as in Equation 1 using the
results from 3 and a set of simulated orebodies; and
5. quantify the risk in the resulting schedule and key project
indicators using simulations of the related orebody.
Stochastic integer programming for mine
production scheduling
Stochastic integer programming (SIP) provides a framework for
optimising mine production scheduling considering uncertainty
(Dimitrakopoulos and Ramazan, 2008). A specific SIP
formulation is briefly shown here that generates the optimal
production schedule using equally probable simulated orebody
models as input, without averaging the related grades. The optimal
production schedule is then the schedule that can produce the
maximum achievable discounted total value from the project,
given the available orebody uncertainty described through a set of
allows the management of geological risk in terms of not meeting
planned targets during actual operation. This is unlike the
traditional scheduling methods that use a single orebody model,
and where risk is randomly distributed between production periods
while there is no control over the magnitude of the risks on the
schedule.
The general form of the objective function is expressed as:
⎡ n ⎤
( )
p m
MAX ∑ ⎢ ∑ E ⎧⎨( NPV) ⎫⎬ b ti − ∑ c tou d tosu + c to1 d tos1 + c tgu d tgsu + c tg1 d tgs1 ⎥,
t
t =1 ⎣ i = 1 ⎩ i
⎭ s=1 ⎦
(2)
where:
p is the total production periods
n is the number of blocks
bi t is the decision variable for when to mine block i (if mined
in period t, bit is 1 and otherwise bit is 0)
The c variables are the unit costs of deviation (represented by
the d variables) from production targets for grades and ore
tonnes. The subscripts u and l correspond to the deviations and
costs from excess production (upper bound) and shortage in
production (lower bound), respectively, while s is the simulated
orebody model number, and g and o are grade and ore production
targets. Figure 6 graphically shows the second term in Equation 2.
Note that the cost parameters in Equation 2 are discounted by
time using the geological risk discount factor developed in

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 15
R DIMITRAKOPOULOS

FIG 5 - Steps needed for the stochastic production scheduling with simulated annealing. S1... Sn are realisations of the orebody grade
through a sequential simulation algorithm. Seq1…Seqn are the mining sequences for each of S1…Sn. Mining rates are input to the process.

EXAMPLES AND VALUE OF THE STOCHASTIC

FRAMEWORK
The example discussed herein shows long-range production
scheduling with both the simulating annealing approach in
Section 3.1 and SIP model in Section 3.2. Section 3.3 focuses on
the topic of stochastically optimal pit limits. The application
used is at a copper deposit comprising 14 480 mining blocks.
The scheduling considers an ore capacity of 7.5 M tonnes per
year and a maximum mining capacity of 28 M tonnes. All results
are compared to the industry’s ‘best practice’: a conventional
schedule using a single estimated orebody model and Whittle’s
approach (Whittle, 1999).

Simulated annealing and production schedules

FIG 6 - Graphic representation of the way the second component of
the objective function in Equation 2 minimises the deviations from
production targets while optimising scheduling. This leads to
schedules where the potential deviations from production targets
are minimised, leading to schedules that seek to mine first not only
for high-grade mining blocks, but also with high probability to be ore.
The results for simulated annealing and the method in Equation 1
are summarised in Figures 7 to 10. The risk profiles for NPV, ore
tonnages, and waste production are respectively shown in
Figures 7, 8, and 9. Figure 10 compares with the equivalent best
conventional practice and reports a difference of 25 per cent in
terms of higher NPV for the stochastic approach.

Dimitrakopoulos and Ramazan (2004). The geological risk

discount rate (GRD) allows the management of risk to be
distributed between periods. If a very high GRD is used, the
lowest risk areas in terms of meeting production targets will be
mined earlier and the most risky parts will be left for later
periods. If a very small GRD or a GRD of zero is used, the risk
will be distributed at a more balanced rate among production
periods depending on the distribution of uncertainty within the
mineralised deposit. The ‘c’ variables in the objective function
(Equation 2) are used to define a risk profile for the production,
and NPV produced is the optimum for the defined risk profile. It
is considered that if the expected deviations from the planned
amount of ore tonnage having planned grade and quality in a
schedule are high in actual mining operations, it is unlikely to
achieve the resultant NPV of the planned schedule. Therefore,
the SIP model contains the minimisation of the deviations
together with the NPV maximisation to generate practical and
feasible schedules and achievable cash flows. For details, please
see Ramazan and Dimitrakopoulos (2008) and Dimitrakopoulos FIG 7 - Risk based life-of-mine production schedule (cumulative net
and Ramazan (2008). present value risk profile).

16 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 STOCHASTIC MINE PLANNING – METHODS, EXAMPLES AND VALUE IN AN UNCERTAIN WORLD

Stochastic integer programming and production

schedules
The application of the SIP model in Equation 2, using pit limits
derived from the conventional optimisation approach, forecasts
an expected NPV at about $238 M. When compared to the
equivalent traditional approach and related forecast, the value of
the stochastic framework is $60 M, or a contribution of about
25 per cent additional NPV to the project. Note that unlike
simulated annealing, the scheduler decides the optimal waste
removal strategy, which is the same as the one used in the
conventional optimisation with which we compare.
Figure 11 shows a cross-section of the two schedules from the
copper deposit: one obtained using the SIP model (bottom) and
the other generated by a traditional method (top) using a single
estimated orebody model. Both schedules shown are the raw
outputs and need to be smoothed to become practical. It is
important to note that:
• the results in the second case study are similar in a
percentage improvement when compared to other stochastic
FIG 8 - Risk based life-of-mine production schedule (ore risk profile). approaches such as simulated annealing; and
• although the schedules compared in the studies herein are not
smoothed out, other existing SIP applications show that the
effect of generating smooth and practical schedules has
marginal impact on the forecasted performance of the related
schedules, thus the order of improvements in SIP schedules
reported here remains.

Stochastically optimal pit limits

The previous comparisons were based on the same pit limits
deemed optimal using best industry practice (Whittle, 1999).
This section focuses on the value of the proposed approaches
with respect to stochastically optimal pit limits. Both methods
described above consider larger pit limits and stop when
discounted cash flows are no longer positive. Figures 12 and 13
show some of the results. The stochastically generated optimal
pit limits contain an additional 15 per cent of tonnage when
compared to the traditional (deterministic) ‘optimal’ pit limits,
add about 10 per cent in NPV to the NPV reported above from
stochastic production scheduling within the conventionally
optimal pit limits, and extend the life-of-mine. These are
substantial differences for a mine of a relatively small size and
short life-of-mine. Further work shows that there are additional
FIG 9 - Risk based life-of-mine production schedule (waste risk improvements on all aspects when a stochastic framework is
profile). used for mine design and production scheduling.

FIG 10 - Net present value of conventional and stochastic (risk based) schedules and corresponding risk profiles.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 17
R DIMITRAKOPOULOS

CONCLUSIONS
Mining Periods Starting from the limits of the current orebody modelling and
1 life-of-mine planning optimisation paradigm, an integrated
2 risk-based framework has been presented. This framework
3 extends the common approaches in order to integrate both
4 stochastic modelling of orebodies and stochastic optimisation in
5 a complementary manner. The main drawback of estimation
6 techniques and traditional approaches to planning is that they are
7
8 unable to account for the in situ spatial variability of the deposit
grades; in fact, conventional optimisers assume perfect knowledge
of the orebody being considered. Ignoring this key source of risk
and uncertainty can lead to unrealistic production expectations as
well as suboptimal mine designs.
The work presented herein shows that the stochastic
framework adds higher value in production schedules in the
order of 25 per cent, and will be achieved regardless of which
method from the two presented is used. Furthermore, stochastic
optimal pit limits are shown to be about 15 per cent larger in
terms of total tonnage, compared to the traditional (deterministic)
FIG 11 - Cross-sectional views of the Stochastic integer optimal pit limits. This difference extends the life-of-mine and
programming (bottom) and traditional schedule (top) for a copper adds approximately ten per cent of net present value (NPV) to the
deposit. NPV reported above from stochastic production scheduling within
the conventionally optimal pit limits.
The new approach yielded an increment of ~30 per cent in the
NPV when compared to the conventional approach. The ACKNOWLEDGEMENTS
differences reported are due to the different scheduling patterns,
the waste mining rate, and an extension of the pit limits which Thanks are in order to the International Association of
yielded an additional ~5.5 thousand tonnes of metal. Mathematical Geosciences for the opportunity to present this

FIG 12 - Life-of-mine cumulative cash flows for the conventional approach, simulated annealing and SIP, compared to results
from conventionally derived optimal pit limits.

FIG 13 - Stochastic pit limits are larger than the conventional ones; physical scheduling differences are expected when bigger pits
are generated.

18 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 STOCHASTIC MINE PLANNING – METHODS, EXAMPLES AND VALUE IN AN UNCERTAIN WORLD
work as their distinguished lecturer. The support of the COSMO
Laboratory and its industry members AngloGold Ashanti,
Barrick, BHP Billiton, De Beers, Newmont, Vale and Vale Inco,
as well as NSERC, the Canada Research Chairs Program and
CFI is gratefully acknowledged. Thanks to R Goodfellow for
editorial assistance.
REFERENCES
Abdel Sabour, S A and Dimitrakopoulos, R, 2010. Accounting for joint
ore supply, metal price and exchange rate uncertainties in mine
design, in Advances in Orebody Modelling and Strategic Mine
Planning I (ed: R Dimitrakopoulos), pp 29-34 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Albor Consquega, F and Dimitrakopoulos, R, 2009. Stochastic mine
design optimization based on simulated annealing: Pit limits,
production schedules, multiple orebody scenarios and sensitivity
analysis, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 118(2):A80-A91.
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
Reserve Estimation, 216 p (Elsevier Science: Amsterdam).
Dimitrakopoulos, R, Farrelly C T and Godoy, M, 2002. Moving forward
from traditional optimization: Grade uncertainty and risk effects in
open-pit design, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 111:A82-A88.
Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty based production
scheduling in open pit mining, SME Transactions, 316:106-112.
Dimitrakopoulos, R, and Ramazan, S, 2008. Stochastic integer
programming for optimizing long term production schedules of open
pit mines: Methods, application and value of stochastic solutions,
Transactions of the Institutions of Mining and Metallurgy, Mining
Technology, 117(4):A155-A167.
Geman, S and Geman, D, 1984. Stochastic relaxation, Gibbs distribution,
and the Bayesian restoration of images, IEEE Transactions on
Pattern Analysis and Machine Intelligence, PAMI-6(6):721-741.
Godoy, M, 2010. A risk analysis based framework for strategic mine
planning and design – Method and application, in Advances in
Orebody Modelling and Strategic Mine Planning I (ed: R
Dimitrakopoulos), pp 21-28 (The Australasian Institute of Mining
and Metallurgy: Melbourne).
Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling, SME Transactions,
316:43-50.
Journel, A G, 1994. Modelling uncertainty: Some conceptual thoughts, in
Geostatistics for the Next Century (ed: R Dimitrakopoulos) (Kluwer
Academic: Dordrecht).
Advances in Orebody Modelling and Strategic Mine Planning I
Kent, M, Peattie, R and Chamberlain, V, 2007. Incorporating grade
uncertainty in the decision to expand the main pit at the Navachab
gold mine, Namibia, through the use of stochastic simulation, in
Orebody Modelling and Strategic Mine Planning, second edition (ed:
R Dimitrakopoulos), pp 207-216 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimization model
for open pit mine planning: Application and risk analysis at a copper
deposit, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 116(3):A109-A118.
Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under
metal uncertainty – Application of stochastic mathematical
programming at an open pit copper mine and comparison to
conventional scheduling, in Advances in Orebody Modelling and
Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Lerchs, H and Grossmann, I F, 1965. Optimum design of open-pit mines,
Transactions of the Canadian Institute of Mining and Metallurgy,
LXVII:47-54.
Meagher, C, Abdel Sabour, S A and Dimitrakopoulos, R, 2010. Pushback
design of open pit mines under geological and market uncertainties,
in Advances in Orebody Modelling and Strategic Mine Planning I
(ed: R Dimitrakopoulos), pp 291-298 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Menabde, M, Froyland, G, Stone, P and Yeates, G, 2007. Mining
schedule optimisation for conditionally simulated orebodies, in
Orebody Modelling and Strategic Mine Planning, second edition (ed:
R Dimitrakopoulos), pp 379-384 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Mustapha, H and Dimitrakopoulos, R, 2010. Geologically enhanced
simulation of complex mineral deposits through high-order spatial
cumulants, in Advances in Orebody Modelling and Strategic Mine
Planning I (ed: R Dimitrakopoulos), pp 309-320 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
long-term production scheduling for open pit mines with a new
integer programming formulation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 385-392 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Ramazan, S and Dimitrakopoulos, R, 2008. Production scheduling with
uncertain supply – A new solution to the open pit mining, COSMO
research report, number 2, pp 257-294.
Whittle, J, 1999. A decade of open pit mine planning and optimisation –
The craft of turning algorithms into packages, in Proceedings
APCOM ’99, Computer Applications in the Minerals Industries:
28 International Symposium, pp 15-24 (Colorado School of Mines:
Golden).
Spectrum Series Volume 17 19

A Risk Analysis Based Framework for Strategic Mine Planning
and Design – Method and Application
M Godoy1
ABSTRACT
Assessment and management of orebody uncertainty is critical to
strategic mine planning. This paper presents an approach that consists of
a series of procedures for risk assessment in pit optimisation and design.
Multiple block grade simulations are processed in Whittle Software to
produce a distribution of possible outcomes in terms of net present value.
Examples from an open pit mine are used to illustrate the practical
application of the methodology.
INTRODUCTION
Traditionally, determination of the spatial distribution of grades
in an orebody model is based on geostatistical estimation. The
main drawback of estimation techniques, be they geostatistical or
not, is that they are unable to reproduce the in situ spatial
variability as inferred from the available data. Ignoring such a
consequential source of risk and uncertainty may lead to
unrealistic production plans (eg Dimitrakopoulos, Farrelly and
Godoy, 2002). In dealing with uncertainty on the spatial
distribution of attributes of a mineral deposit, several models of
the deposit can be generated based on and conditional to the
same available data and their statistical characteristics. These
models are all constrained to:
• reproduce all available information and their statistics, and
• represent equally probable models of the actual spatial
distribution of grades.
The availability of multiple equally probable models of a mine
deposit enables mine planners to assess the sensitivity of pit
design and long-term production scheduling to geological
uncertainty. This approach has been proposed by many authors
over the last 20 years (David, 1988; Journel, 1992; Ravenscroft,
1992; Dimitrakopoulos, 1998; Kent, Peattie and Chamberlain,
2007; and others). Figure 1 illustrates the difference between the
traditional process used to convert a mineral resource into an ore
reserve and the risk based approach based on the technique of
conditional simulation.
The goal of this paper is to provide mining planning engineers
a series of procedures that can be used to consider the effects of
grade uncertainty in mine planning studies. Four cases have been
selected to illustrate different types of applications:
1. uncertainty analysis of an ultimate pit shell – Net Value
(NV), costs, tonnage, grade and metal;
2. identification of areas of upside potential and downside
risk;
3. trade-off analysis for cut-back depletion strategies; and
4. assessment of uncertainty related to ore blocks driving the
increment between the successive pit shells.
In the following sections, each application is developed
separately and includes a step-by-step description of the
procedure and a discussion of the results obtained in a real case
application.
1. MAusIMM, Expert Mining Engineer/Geostatistician, Golder
Associates, Av 11 de Septiembre 2353, Piso 2, Proviendencia
Santiago, Chile. Email: mgodoy@golder.cl
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
UNCERTAINTY ANALYSIS FOR AN ULTIMATE
PIT SHELL
The objective of this analysis is to evaluate the sensitivity of key
pit optimisation results to grade uncertainty. The process known as
pit optimisation is traditionally carried out based on the estimated
resource model and using the nested pit implementation of the
Lerchs-Grossmann algorithm of the Whittle Software (Lerchs and
Grossmann, 1965; Whittle, 1999). The result of the optimisation
process is a series of incremental pit shells. Different criteria can
be used to select the ultimate pit shell including net pit value and
the net present value based on a referential mining sequence. This
ultimate pit shell is then used as the basis for pit design and
planning. Using conditionally simulated models as input, Whittle’s
analysis program (FDAN) may be used allowing the quantitative
assessment of risk due to uncertainty on the real, but unknown,
distribution of grades. In the procedure proposed below an
ultimate pit shell produced in Whittle is evaluated against a series
of simulated models of the orebody.
To assess uncertainty on the main parameters driving the
selection of the ultimate pit shell the following procedure is
proposed:
• From the Whittle result file produced by the pit optimisation
process, generate a Whittle pit list file containing information
about the smallest numbered pit that each block is part of.
• Apply the pit list file produced in Step 1 to each one of the
simulated orebody models.
• Run the analysis program configured to generate the same
information, previously generated by analysis on the original
pit optimisation. The analysis must be carried out for each
one of the available simulated models.
The above procedure generates a range of alternative outcomes
for the original optimisation process. This allows the planner to
evaluate the likely range of contained ore, metal and a series
of key economic indicators. Figure 2 shows the predicted ore
tonnage produced by risk analysis. Up until pit 27, the cloud of
cumulative tonnages derived from the simulated models present
an average decrease of approximately 9.76 per cent in relation to
the tonnage predicted by the estimated model. At pit 27 the
estimated model indicates approximately 180.3 Mt against an
expected value of 164.6 Mt derived from the simulations. The
expected outcomes of contained ore go from 163.3 Mt to
166.1 Mt, which corresponds to a range of -0.83 per cent to
+0.92 per cent in relation to the expected value derived from the
simulations. The same type of analysis carried out on the average
mill feed grade (Figure 3) shows a decreasing overestimation in
grades of the estimated model in relation to the simulated
models. At pit shell number 1, the estimated model predicts an
average mill feed grade of 2.6 g/t, while the simulations indicate
an expected grade of 2.2 g/t. At pit 27 the estimated model
indicates approximately 1.98 g/t against an expected value of
1.95 g/t derived from the simulations. The risk profile on mill
feed grade goes from 1.93 g/t to 1.98 g/t, which corresponds to a
range of -1.24 per cent to +1.32 per cent in relation to the
expected value. Figure 4 shows the predicted recovered metal
produced by the analysis. Up until pit 27, the cloud of cumulative
recovered ounces derived from the simulated models present an
average decrease of approximately 17.41 per cent in relation to
the ounces predicted by the estimated model. This result
indicates that the estimated model is potentially overestimating
Spectrum Series Volume 17 21
M GODOY

1
Single,
Unknown, often

Probability
true precise,
answer wrong
Single estimated
Traditional view answer
model

Reserves
Mining Process or
Transfer Function
Financial and
Mine Design &
Orebody Model Production
Production Scheduling
Forecasts

Multiple probable Accurate

1
models uncertainty
estimation

Probability
Risk oriented view

Reserves

FIG 1 - Conversion of a mineral resource into an ore reserve, traditional and risk oriented views.

Ore Quantity
350 Recovered Metal
18000
300
16000
250
14000
Ore (Mt)

200 12000
Gold (Moz)

ResSim1 ResSim2 10000

150 ResSim3 ResSim4
ResSim5 ResSim6 8000 ResSim1 ResSim2
ResSim7 ResSim8 ResSim3 ResSim4
100
ResSim9 ResSim10 6000 ResSim5 ResSim6
ResSim7 ResSim8
ResSim11 ResSim12
4000 ResSim9 ResSim10
50 ResSim13 ResSim14
ResSim11 ResSim12
ResSim15 EstMod
2000 ResSim13 ResSim14
0 ResSim15 EstMod

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
Pit Number
Pit Number

FIG 2 - Uncertainty in ore tonnes for incremental pit shells.

FIG 4 - Uncertainty in recovered ounces for incremental pit shells.
Mill Feed Grade
3
ResSim1 ResSim2
2.8 ResSim3 ResSim4 Undiscounted Cash Flow (Net Value)
ResSim5 ResSim6
2.6 ResSim7 ResSim8
1400
ResSim9 ResSim10
2.4 ResSim11 ResSim12
ResSim13 ResSim14 1200
Grade (g/t)

2.2 ResSim15 EstMod

2 1000
Net Value (M $)

1.8 800
1.6
600
1.4

1.2 400 ResSim1 ResSim2 ResSim3 ResSim4

ResSim5 ResSim6 ResSim7 ResSim8
1 ResSim9 ResSim10 ResSim11 ResSim12
200
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 ResSim13 ResSim14 ResSim15 EstMod

Pit Number 0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51

FIG 3 - Uncertainty in mill feed grade for incremental pit shells. Pit Number

grades as an effect of excessive smoothing. At pit 27 the
estimated model indicates 9.84 Moz against an expected value of
8.85 Moz derived from the simulations. The risk profile on
recovered ounces goes from 8.72 Moz to 8.98 Moz, which
corresponds to a range of -1.4 per cent to +1.58 per cent in
relation to the expected value.
The combination of the overestimation in ore tonnage and mill
feed grade has a direct impact on the performance of the pit by
pit Net Value. Figure 5 presents the results obtained for the pit
Net Value.
It suggests that the estimated model overestimates the pit value
over all optimised pit shells. Up until pit 27, the cloud of
cumulative Net Value derived from the simulated models present
FIG 5 - Uncertainty in Net Value for incremental pit shells.
an average decrease of approximately 33.39 per cent in relation
to the Net Value predicted by the estimated model. It also
indicates a downside potential for pit 27 with expected NV
equivalent to $951.6 M or a 26.12 per cent decrease in relation to
the $1288 M obtained for the analysis on the estimated block
model. It is interesting to note that if Net Value was to be used as
the criterion for the selection of the ultimate pit shell the
simulations would agree with the estimated model by indicating
pit 27. The range of expected Net Values goes from $890 M to
$1013 M, which corresponds to -6.47 per cent to +6.41 per cent
in relation to the expected value.

22 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 A RISK ANALYSIS BASED FRAMEWORK FOR STRATEGIC MINE PLANNING AND DESIGN – METHOD AND APPLICATION

IDENTIFICATION OF UPSIDE AND DOWNSIDE
POTENTIAL
The goal of this analysis is to explore the possible downside/
upside potential of the selected ultimate pit shell regarding the
available grade uncertainty models. To achieve this independent
pit optimisation runs are carried out on each simulated model.
The analysis is divided into two parts:
1. First, each optimisation output is evaluated in terms of
contained ore, grade, metal and pit value. This provides a
quantification of the project potential given realistic
scenarios of the spatial distribution of grades.
2. The second part looks into the spatial extends of a specific
pit shell as produced for each independent optimisation.
The comparison of this ‘cloud’ of pit shells against a given
pit design provides an assessment of areas with
upside/downside potential and can be used to define targets
for additional drilling. It also provides an indication on the
robustness of a given pit design in relation to grade
uncertainty.
The procedure proposed for the development of this analysis is
the following:
• Create a new project in Whittle and import each one of the
simulated models. Next, set the optimisation parameters for
the optimisation run.
• Apply the same parameters to each model and run the
optimisation. The process will generate a series of Whittle
result files, one for each simulated model. This step usually
requires a large amount of disk space.
• Run analysis program configured to generate the relevant
summary information. The analysis must be carried out for
each one of the available simulated models.
• Produce cross-sections for a selected pit number over all
estimated and simulated pits present an equivalent quantity of
contained rock, that is, they have approximately the same
volume. From around pit shell 25, there is a clear separation with
the simulated pits showing a progressive increase in comparison
to the estimated pit. This scenario remains the same until pit
shell 39, when the estimated pit starts to converge to the cloud of
simulated pits. Pit shell 27 is of particular interest as it
corresponds to the pit selected as a basis for pit design. For pit
shell 27, the average contained rock over the simulated pits is
925 Mt against 828 Mt in the estimated pit, which corresponds to
an increase of approximately 11.7 per cent. Figure 7 presents the
results obtained for the total contained ore. In this case,
estimated and simulated pits show similar behaviour as the total
rock, only that the magnitude of the differences is smaller. For pit
shell 27, the average contained ore over the simulated pits
is 183.9 Mt against 180.3 Mt in the estimated pit, which
corresponds to an increase of approximately two per cent. The
risk profile on the contained ore goes from 174 Mt to 195 Mt,
which corresponds to a range of -5.92 per cent to +6.06 per cent
in relation to the expected value. In terms of mill feed grade, the
estimated pit starts with 2.61 g/t against an average 2.43 g/t over
the simulated pits. This difference decreases with incremental pit
shells and become equivalent at pit shell 27 (Figure 8).
The risk profile on mill feed grade goes from 1.95 g/t to
2.02 g/t, which corresponds to a range of -1.61 per cent to
+1.92 per cent in relation to the expected value. Figure 9 and
Figure 10 present the results obtained for recovered metal and
total pit value.
As expected, the recovered metal for estimated and simulated
pits follow the same trends seen on the ore tonnage graphs. For
pit shell 27, the recovered gold over simulated pits is 10.03 Moz
Ore Quantity
350
300

optimised models. 250

Contrary to developing a risk analysis on a given ultimate pit

Ore (Mt)

200
shells, as carried out in the previous section, the above procedure ResSim1 ResSim2
150
generates alternative sets of incremental shells, one set for each ResSim3
ResSim5
ResSim4
ResSim6
simulated model. Figure 6 shows the total rock contained on 100 ResSim7 ResSim8
ResSim9 ResSim10
each incremental shell as produced by each independent ResSim11 ResSim12
50 ResSim13 ResSim14
optimisation. ResSim15 EstMod

The thick orange line corresponds to the results obtained from 0

the analysis on the incremental shells optimised on the estimated
model, hereafter termed estimated pit. The thin lines correspond
to the results obtained from the analysis on the incremental shells
optimised on each one of the simulated models, hereafter termed
simulated pits. The figure shows that up to pit shell 24 the
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
Pit Number
FIG 7 - Total ore tonnage contained on incremental pit shells.
Estimated pit is depicted by the thick orange line and the
remaining coloured thin lines correspond to simulated pits.

Total Movement
2500 Gold Grade
3
ResSim1 ResSim2 ResSim1 ResSim2
ResSim3 ResSim4 2.8 ResSim3 ResSim4
ResSim5 ResSim6 ResSim5 ResSim6
2000 ResSim7 ResSim8
ResSim7 ResSim8 2.6
ResSim9 ResSim10 ResSim9 ResSim10

ResSim11 ResSim12 2.4 ResSim11 ResSim12

1500 ResSim13 ResSim14
Rock (Mt)

ResSim13 ResSim14
Grade (g/t)

2.2 ResSim15 EstMod

ResSim15 EstMod
2
1000
1.8

1.6
500
1.4

1.2
0
1
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
Pit Number
Pit Number

FIG 6 - Total rock tonnage contained on incremental pit shells. FIG 8 - Average grade contained on incremental pit shells.
Estimated pit is depicted by the thick orange line and the Estimated pit is depicted by the thick orange line and the
remaining coloured thin lines correspond to simulated pits. remaining coloured thin lines correspond to simulated pits.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 23
M GODOY

Recovered Gold
350

300

250
Ore (Mt)

200

ResSim1 ResSim2
150 ResSim3 ResSim4
ResSim5 ResSim6
100 ResSim7 ResSim8
ResSim9 ResSim10
ResSim11 ResSim12
50 ResSim13 ResSim14
ResSim15 EstMod

0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51
Pit Number

FIG 9 - Total recovered metal contained on incremental pit shells.

Estimated pit is depicted by the thick orange line and the
remaining coloured thin lines correspond to simulated pits.

Undiscounted Cash Flow (Net Value)

1400

1200

1000
Net Value (M$)

800

ResSim1 ResSim2
600
ResSim3 ResSim4
ResSim5 ResSim6
400 ResSim7 ResSim8
ResSim9 ResSim10
ResSim11 ResSim12
200 ResSim13 ResSim14
ResSim15 EstMod
0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51

Pit Number

FIG 10 - Total pit value of incremental pit shells. Estimated pit is

depicted by the thick orange line and the remaining coloured thin
lines correspond to simulated pits.

against 9.84 MOz in the estimated pit, which is an increase of

approximately 0.9 per cent. This combination of slightly higher
metal with considerable more rock tonnage in the simulated pits
translates into a reduced net pit value when compared to the
estimated pit. The estimated pit presents a consistently higher
Net Value until pit shell 39 when the total rock tonnage becomes
equivalent to that shown on the simulated pits. For pit shell 27
the average Net Value over simulated pits is 1159 million dollars
against 1288 million in the estimated pit (approximately ten per
cent less). The risk profile on Net Value goes from $1068 M to
$1292 M, which corresponds to a range of -7.86 per cent to FIG 11 - Cross-sections produced for pit shell 27 over all
+6 per cent in relation to the expected value. It is interesting to
optimised models overlaid with actual pit design.
note that as for the estimated model the maximum Net Value
over all simulated pit corresponds to pit shell number 26. This
shows that pit shell 26 is quite robust with regards to grade extension represents a major downside potential zone and goes
uncertainty. Another conclusion that can be drawn from these from the actual pit surface to the lowest levels of the pit. The
results, which is coherent with the results obtained in the analysis simulated pits indicate an upside potential region at the
presented in the previous section, is that there is a global southwest zone of the pit where the simulated pits reach levels
overestimation of grades and ore tonnage by the estimated model that are deeper than the current pit design. The spread of cloud of
as compared to the simulated models. simulated pits is shown to be low from the surface down to level
A series of cross-sections were produced for pit shell 27 over -350. Bellow that the spread increases considerably. This reflects
all optimised models and for a pit design. These cross-sections the increasing uncertainty on the distribution of grades at depth
were overlaid and are presented in Figure 11. The main and is directed related to the lack of drilling.
conclusions drawn from the analysis of these cross-sections are
the following. The simulated pits closely honour the eastern wall TRADE-OFF ANALYSIS
of the current pit design, showing that the eastern slope is stable
with relation to grade uncertainty. In general, there is a more The goal of this analysis is to quantify the impact of grade
pronounced fluctuation in the western wall which indicates uncertainty to tonnage, grades, metal and Net Value of two
higher levels of grade uncertainty. The current design has an different mining strategies. Scenario ‘A’ considers the depletion
extension of the western wall, which is not included in of a cut back as a single stage, while scenario ‘B’ defines two
the optimisations of both estimated and simulated pits. This separate stages. The analysis consists on the quantification of

24 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 A RISK ANALYSIS BASED FRAMEWORK FOR STRATEGIC MINE PLANNING AND DESIGN – METHOD AND APPLICATION

uncertainty on key mining physicals and economic parameters

for the two mining scenarios considered. The objective is to
300
evaluate if one of the scenarios is any better in terms of the
compromise between Net Value and risk exposure. The Maximum

Ore Quantity (Mt)

a
250
Minimum
procedure proposed to develop the analysis is the following:
Expected
1. generate a wireframe describing the cut-back, 200 Estimated

2. filter the block model against the wireframe and retain the 150
blocks lying inside the cut-back as new block model,
100
3. export the block model produced above into a Whittle
Model File,
50
4. repeat steps 2 and 3 for each one of the simulated models, OneStage FirstStage SecondStage

5. for each model produced in the previous steps calculate the

relevant summary information, and FIG 13 - Risk profiles on contained ore of two mining scenarios.

6. repeat steps 1 to 6 for stages 1 and 2 that correspond to

another mining scenario.
The procedure was carried out for a total of 16 models,
corresponding to 15 grade simulations plus the estimated model. 1.8 Maximum

Metal Quantity (Moz) a

Minimum
Figure 12 presents a 3D view of the cut-back showing its 1.5
Expected
position in relation to the current pit design. Figure 13 shows the 1.3 Estimated
results in terms of contained ore for each mining scenario:
1.0
• the first profile corresponds to the scenario ‘A’, which 0.8
corresponds to the cut-back depletions as a single stage; and
0.5
• the other profiles correspond to the first and second stages of
0.3
scenario ‘B’.
0.0
OneStage FirstStage SecondStage

FIG 14 - Risk profiles on recovered metal of two mining scenarios.

cent and 3.61 per cent respectively. Figure 15 summarises the

FIG 12 - View of the cut-back aganst the pit design. The colours
indicate different incremental pit shells.
The expected combined ore tonnage of scenario ‘B’ is 2.3 per
cent higher than scenario ‘A’ (28.2 Mt against 27.6 Mt). The risk
profile for scenario ‘A’ shows a range of variation that
corresponds to 5.2 per cent of the expected ore tonnage. For the
two stages of scenario ‘B’, these ranges correspond to 5.19 per
cent and 6.51 per cent respectively. For scenario ‘A’, the
contained ore predicted by the estimated model is 4.16 per cent
higher than the expected tonnage derived from the simulations.
For the first and second stages, this difference corresponds to
results in terms of Net Value.
The expected Net Value of scenario ‘B’ is approximately 22.3 per
cent lower than in scenario ‘A’ ($76.5 M against $98.4 M). The
risk profile for scenario ‘A’ shows a range of variation that
corresponds to 50 per cent of the expected Net Value. For
scenario ‘B’ the ranges correspond to 68.61 per cent and
141.2 per cent for the first and second stages respectively. The
combined expected range of variation for scenario ‘B’
corresponds to 67.4 per cent. For scenario ‘A’ the Net Value
predicted by the analysis on the estimated model is 76.85 per
cent higher than the expected tonnage derived from the
simulations. For the first and second stages this difference
corresponds to 162.47 per cent and -2.48 per cent respectively.
The results obtained in terms of the risk profiles indicate that
both scenarios present high risk of not achieving predicted Net
Value. In addition, it was clearly identified that the volume
related to stage 1 of the scenario ‘B’ is the one with all the risk.
200
Maximum
175
Minimum
Net Value ($M) a

150
4.63 per cent and 9.89 per cent respectively. Figure 14 shows the Expected
Estimated
results in terms of recovered gold. 125

The expected recovered metal of scenario ‘B’ is approximately 100

1.7 per cent higher than in scenario ‘A’ (1.63 Moz against 75
1.60 Mt). The risk profile for scenario ‘A’ shows a range of 50
variation that corresponds to 7.1 per cent of the expected 25
tonnage. For scenario ‘B’ this ranges correspond to 8.5 per cent
0
and 12.3 per cent. For scenario ‘A’ the recovered gold predicted
OneStage FirstStage SecondStage
by the estimated model is 10.43 per cent higher than the expected
tonnage derived from the simulations. For the first and second
stages of scenario ‘B’, this difference corresponds to 16.70 per FIG 15 - Risk profiles on pit value of two mining scenarios.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 25
M GODOY

It is important to note that this analysis is rather simplistic in the
sense that the time effect of money is not included. Ideally, a
mining schedule should be developed in order to account for the
mine sequencing. However, the second scenario roughly
accounts for the sequencing by developing the depletion in two
stages. The analysis indicated that the risk of missing the target
when mining the volume related to the first stage of the second
scenario is extremely high. This warrants a detailed review of the
estimated grades in this volume.
RISK ANALYSIS ON ORE BLOCKS DRIVING
A PIT INCREMENT
The aim of this analysis is the quantification of uncertainty on
the ore blocks driving the increment between two successive
Whittle pit shells. To assess uncertainty on the main parameters
driving increment between two successive pit shells, the follow
procedure is proposed:
1. From the Whittle result file produced by the pit
optimisation process generate a Whittle pit list file
containing information about the smallest numbered pit that
each block is part of.
2. Use the re-blocking program – apply the pit list file
produced in step 1 to each one of the simulated block
models. This will create a set of results files.
3. Produce cross-sections for the incremental pit shells.
4. Generate the summary pit information for the two pits. The
analysis must be carried out for each one of the available
simulated models. Derive the information referent to the
incremental volume by subtracting the cumulative mining
physicals (ore and metal quantities) and economic values
(Net Value, processing cost and mining cost) between the
two successive pit shells.
The above procedure is similar to the procedure used in the
first section of this paper. The difference is that here the analysis
is limited to two specific pit shells. The increment from pit 26 to
pit 27 contains approximately 53 Mt of rock and is located in the
southern end of the pit. Table 1 presents the results obtained by
the analysis of the pit increment.
TABLE 1
Risk profile for mining physicals and economic parameters for
the increment between pits 26 and 27.
Expected Minimum Maximum Range
Ore (×1000t) 5925 5706 6222 516
Metal (oz) 341 739 309 451 362 982 53 531
Grade (g/t) 1.95 1.93 1.98 0.05
Mcost (×1000$) -103 594 -103 716 -103 521 195
Pcost (×1000$) -70 783 -74 285 -68 238 6047
Net value (×1000$) 7381 -8164 17 685 25 849
The risk profile on the contained ore shows a range between
5.7 Mt and 6.2 Mt, with an expected tonnage of approximately
5.9 Mt. Low ranges of variation are also shown for the risk
profiles on grade, metal content and costs. The main issue here is
in relation to the Net Value, which has a chance of being
negative. Its risk profile goes from approximately $-8.2 M to
$17.7 M. The reason for the increase in the risk profile the
mining cost associated to a high stripping ratio (~9), which
makes the Net Value oversensitive to possible variations on the
recovered gold. It is important to notice that the risk profile
indicates an expected Net Value for this increment of $7.4 M. In
fact, only one out of 15 simulated models presented a negative
value for the Net Value. The results indicate a relatively low
uncertainty in tonnages and grades contained between pits 26
and 27. However, this relatively low uncertainty becomes a
critical issue due the high stripping ratio which makes the
increment’s Net Value very sensitive to grade uncertainty as well
as gold price.
Several cross-sections have been generated to show the pit
region relative to the increment between pits 26 and 27. These
sections show the incremental shells, the current pit design and
ore blocks contained inside the increment. The main conclusions
drawn from the analysis of these cross-sections have four
components. First, the major difference between the incremental
Whittle pit shells $550/oz and $560/oz corresponds to a region
located at the southern end of the pit. Most of the incremental ore
blocks have an expected value inside the range of 1.5 g/t to
2.5 g/t. Moreover, most of the incremental ore blocks have more
than 60 per cent chance of being above the cut-off. The
increment contains a high quantity of waste and the pit design
has considerably more waste than pit shell 27.
CONCLUSIONS
The goal of this work was to illustrate different applications of
risk analysis on the effects of grade uncertainty to various
aspects of pit optimisation and design. Four cases have been
carried out to illustrate different types of applications:
1. The first case consisted of an uncertainty analysis on pit
optimisation results – Net Value, tonnage, grade and metal.
The procedure consisted in applying a set of incremental pit
shells, as produced by the pit optimisation process, to a set of
simulated resource models. The subsequent analysis on each
model produced a set of equally probably outcomes for the
mining physicals and economic forecasts given the initial set
of incremental pit shells.
2. The second case identified areas were grade uncertainty has
major impact to the definition of the ultimate pit limits
(upside/downside potential). Rather than developing a risk
analysis on a given set of incremental pit shells, this
procedure consisted in the generation of alternative sets of
incremental shells, one set for each simulated model.
3. The third case aimed at quantifying the impact of grade
uncertainty to tonnage, grades, metal and Net Value of two
different mining scenarios for a given cut-back. The main
objective was to evaluate if one of the scenarios was any
better in terms of the compromise between Net Value and
risk exposure.
4. The fourth case consisted of a risk analysis related to pit
increments. The objective of this analysis is the
quantification of uncertainty on the ore blocks driving the
increment between two successive Whittle pit shells.
The results were presented in two steps:
1. first, each optimisation output was evaluated in terms of
contained ore, grade, metal and pit value; and
2. the second step of the analysis consisted on the generation
of a series of cross-sections.
These cross-sections were taken over all optimised models and
included the actual pit design. Several conclusions have been
drawn from these graphs indicating areas of upside and downside
potential.
This paper presented a set of procedures that enable mine
planning engineers to carry out a series of analysis, which can be
used to evaluate the sensitivity of incremental pit shells and pit
designs to grade uncertainty. The results obtained from the
analysis have shown to provide valuable information, which can
be used to develop mining strategies that are risk resilient in
relation to grade uncertainty.

26 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 A RISK ANALYSIS BASED FRAMEWORK FOR STRATEGIC MINE PLANNING AND DESIGN – METHOD AND APPLICATION
REFERENCES
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
Reserve Estimation, 216 p (Elsevier Science Publishers: Amsterdam).
Dimitrakopoulos, R, 1998. Conditional simulation algorithms for
modelling orebody uncertainty in open pit optimisation,
International Journal of Surface Mining, Reclamation and
Environment, 12:173-179.
Dimitrakopoulos, R, Farrelly, C T and Godoy, M, 2002. Moving forward
from traditional optimisation: grade uncertainty and risk effects in
open-pit design, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 111:A82-A88.
Dimitrakopoulos, R, Martinez, L and Ramazan, S, 2007. Optimising open
pit design with simulated orebodies and Whittle Four-X: A
maximum upside/minimum downside approach, in Orebody
Modelling and Strategic Mine Planning, second edition (ed: R
Dimitrakopoulos), pp 201-206 (The Australasian Institute of Mining
and Metallurgy: Melbourne).
Godoy, M C, 2003. The effective management of geological risk in
long-term production scheduling of open pit mines, PhD thesis,
256 p, The University of Queensland, Brisbane.
Journel, A G, 1992. Computer imaging in the minerals industry – Beyond
mere aesthetics, in Proceedings 23rd APCOM (International
Symposium on Application of Computers and Operations Research in
the Mineral Industry), pp 3-13 (Society for Mining, Metallurgy and
Exploration: Littleton).
Advances in Orebody Modelling and Strategic Mine Planning I
Kent, M, Peattie, R and Chamberlain, V, 2007. Incorporating grade
uncertainty in the decision to expand the main pit at the Navachab
gold mine, Namibia, through the use of stochastic simulation, in
Orebody Modelling and Strategic Mine Planning, second edition
(ed: R Dimitrakopoulos), pp 207-216 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Lerchs, H and Grossmann, I F, 1965. Optimum design of open pit mines,
CIM Bulletin, Canadian Institute of Mining and Metallurgy, volume
58, January.
Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
simulation, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 101:A101-A108.
Whittle, J, 1999. A decade of open pit mine planning and optimisation –
The craft of turning algorithms into packages, in Proceedings
APCOM ’99 International Symposium on Application of Computers
and Operations Research in the Mineral Industry), pp 15-24
(Colorado School of Mines: Colorado).
Spectrum Series Volume 17 27

Accounting for Joint Ore Supply, Metal Price and Exchange Rate
Uncertainties in Mine Design
S A Abdel Sabour1 and R Dimitrakopoulos2
ABSTRACT
Conventional methods for selecting a production plan for an open pit
mine are based on the assumption that ore grades throughout mine life as
well as market variables are known with certainty. Another important
issue that is also being ignored in the conventional procedure is the value
of management flexibility to revise operating policy with time according
to the arrival of new information. Based on this oversimplification of
reality, an optimum mine plan that is thought to be maximising net
present value (NPV) of the project is selected. Given the multiple sources
of risk affecting the estimated NPV, there is a high probability that the
actual value will be different from that estimated. This could mislead
decision making process regarding selecting an optimum mine plan.
This paper proposes a multi-criteria ranking system for selection
between alternative mine designs under both geological and market
uncertainties. The system is based on integrating multiple market and
geological uncertainties as well as the operating flexibility to revise the
ultimate pit limits using a Monte Carlo-based real options valuation
(ROV) model. To compare the methods, the article applies the proposed
system, along with the conventional NPV-based method, to rank possible
mine designs at a Canadian copper mine. It has been found that,
considering multiple risk analysis measures while integrating uncertainty
and operating flexibility results in improving the mine design selection
process.
INTRODUCTION
The ultimate goal in open pit mine planning is to define an
‘optimum’ mine design and a life-of-mine (LOM) sequence of
production. In this respect, the conventional method for open pit
planning starts with modelling the orebody based on the borehole
data and geological information. Then, the mining field is
divided into blocks of regular volume. Based on a deterministic
metal price, each block is assigned a value equal to the gross
value of its metal content minus the applicable production,
processing and refining costs. The optimum production plan is
determined by applying different heuristic and optimisation
algorithms (Whittle, 1988). Commonly, there may be alternative
technically feasible mine plans available that meet operational
and technical constraints. The selection among those plans is
then based on economic reasons. This is carried out by
evaluating each of the possible mine plans and comparing their
economic attractiveness so as to select the most economically
appealing one. The basic assumption in most previous work is
that both the grades of blocks and metal prices are known with
certainty (see for example Whittle, 1988; Ramazan, 2007).
Nevertheless, the real situation in mining industry is much more
complex than this simple assumption suggests. In practice, mine
planners cannot know with certainty the quantity and quality of
ore in the ground. Also, both future metal prices and foreign
exchange rates are impossible to know with certainty.
1. COSMO – Stochastic Mine Planning Laboratory, Department of
Mining and Materials Engineering, McGill University, FDA
Building, 3450 University Street, Montreal QC H3A 2A7, Canada.
Email: sabry.abdelhafezabdelsabour@mcgill.ca
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montreal QC H3A 2A7, Canada.
Email: roussos.dimitrakopoulos@mcgill.ca
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
Among other sources, the geological and market uncertainties
are major sources of risk in mining businesses. An example
for geological uncertainty is illustrated in Figure 1, showing 50
simulations for the grade of an Australian gold mine. It is obvious
that the possible average mill feed grade is highly uncertain.
Ignoring such uncertainty about ore tonnage and grade could
result in substantial losses especially when considering the
capital-intensive nature of mining investments. As reported by
Vallee (2000), a World Bank study conducted in the 1990s showed
that a total loss of $US1106 million resulted from ore reserve
problems in failed mining projects in Canada and the USA.
2.10
2.05
2.00
Mill Feed Grade, g/t Au
1.95
1.90
1.85
1.80
1.75
1.70
1.65
1.60
Possible Average Mill Feed Grade
FIG 1 - Uncertainty of ore grade (modified from Dimitrakopoulos,
Farrelly and Godoy, 2002).
The second major source of risk affecting mine project
profitability is related to the uncertainty about market behaviour
of metal prices and exchange rates. Figure 2 shows, for example,
the average annual market prices of gold and copper over the
period 1980 - 2006 in current US dollars. It is obvious that these
prices are highly volatile and do not keep a constant trend, which
makes it speculative to define deterministic forecasts for future
metal prices. Exchange rate is also another contributor to project
risk. Figure 3 shows how the exchange rates of the currencies of
two major metal exporting countries like Australia and Canada
are uncertain. Therefore, it is difficult for mine planners to have
precise forecasts for these key variables over the mine life.
Under the uncertain geological and market conditions, the
process of selecting a mine plan among different alternatives is
not an easy task. Conventional static financial evaluation
methods based on deterministic geological and market variables
are not well suited to handle the multiple sources of risk and the
dynamic, proactive, nature of management decisions. There is a
need for a more efficient system for mine plan selection under
multiple uncertainties that minimises the need for subjective
judgements. Such system should be based on sound
quantification and management of both the geological and
market uncertainty. Integrating the geological uncertainty into
open pit mine planning was first introduced by Dimitrakopoulos,
Farrelly and Godoy (2002). Godoy and Dimitrakopoulos (2004)
and Leite and Dimitrakopoulos (2007) presented a risk-based
optimisation approach for long-term open pit mine planning
Spectrum Series Volume 17 29
S A ABDEL SABOUR and R DIMITRAKOPOULOS

700 3.40 This paper builds on the work of Dimitrakopoulos, Martinez

Gold price and Ramazan (2007) by quantifying and integrating the market
Copper price 3.00 uncertainty related to metal prices and exchange rates into mine
600 planning. In this respect, the article aims to develop a system for

Copper price, US$/lb

mine plan selection based on multiple value statistics and cash
Gold price, US$/oz

2.60
flows characteristics incorporating the value of management
500
2.20 flexibility to react to the new information. In the next sections,
the proposed selection system will be briefly outlined. Then, it
1.80 will be applied, along with the conventional methods of
400
selection, to select the optimum production schedule for a
1.40
Canadian copper mine. Finally, an investigation of the usefulness
300 of the developed system will be provided.
1.00

200 0.60 A SYSTEM FOR MINE DESIGN SELECTION

1980 1985 1990 1995 2000 2005 UNDER UNCERTAINTY
Year
First, it is worth noting that the different alternative mine designs
FIG 2 - Gold and copper prices uncertainty. could result from different available production scenarios,
different LOM sequences, and so on. In this study, the procedure
described in Dimitrakopoulos, Martinez and Ramazan (2007) for
2.2 generating different mine designs based on simulating multiple
US$/AU$
orebody realisations is used. The mine design selection system,
2 US$/CAN$ or ranking system, proposed in this work takes into account
multiple sources of uncertainty simultaneously and integrates the
operating flexibility to revise the ultimate pit limits based on the
Exchange rate

1.8
new information. The system is outlined in Figure 4 and consists
of three main steps: uncertainty quantification, design valuation
1.6
and design ranking.

1.4
Quantifying geological
1.2 uncertainty

1
Jan-1990 Jan-1993 Jan-1996 Jan-1999 Jan-2002 Jan-2005 Quantifying exchange rate Quantifying metal prices
Time uncertainty uncertainty

FIG 3 - Exchange rates uncertainty.

RO valuation using Monte
considering geological uncertainty. Dimitrakopoulos, Martinez Carlo
and Ramazan (2007) proposed an approach for designing open
pit mines under geological uncertainty based on maximising
upside potential and minimising downside risk.
Design ranking
The other significant source of uncertainty that should be
considered is the market uncertainty. In this respect, the real
options valuation (ROV) can provide a promising tool for better
evaluating alternative production schedules under uncertainty. FIG 4 - Outline of the proposed mine design selection system.
The main advantage of the ROV over the conventional financial
analysis methods based on the discounted cash flow analysis lies Uncertainty quantification
in its ability to evaluate the management flexibility to react to the
new information (Samis and Poulin, 1998 and Monkhouse and In this step, both the geological and market uncertainties are
Yeates, 2005). For more details on real options valuation see, for quantified. First, the geological uncertainty is explored by
example, Brennan and Schwartz (1985), Dixit and Pindyck simulating multiple orebody realisations based on the borehole
(1994), Tufano and Moel (1999), Slade (2001), Kamrad and data using conditional simulation (see Dimitrakopoulos, 2007).
Ernst (2001), Moel and Tufano (2002), Samis et al (2003; 2006). Market uncertainty about metal prices and foreign exchange rates
It is worth mentioning here that both the conventional net present is quantified using also stochastic modelling based on historical
value (NPV) and the ROV calculate the discounted, net present market behaviour. Examples of stochastic models include
value of future cash flows. The two techniques differ in the way Geometric Brownian Motion (GBM) model in Equation 1 (Dixit
they apply discounting and in the way they deal with operating and Pindyck, 1994) and the mean-reversion model in Equation 2
flexibility. Conventional NPV applies a single discount rate to all (Schwartz, 1997):
cash flow components and does not handle flexibility. ROV
allows for differential discounting based on specific risk of each dP
cash flow component and can capture the value of operating = αdt + σdz (1)
flexibility. Throughout this paper the abbreviation ‘NPV’ will be P
used to indicate the conventional, single discount rate, no
dP
flexibility, NPV and the abbreviation ‘ROV’ will be used to = κ(μ − 1nP)dt + σdz (2)
indicate the advanced, multiple-rates, flexibility-based NPV. P

30 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 ACCOUNTING FOR JOINT ORE SUPPLY, METAL PRICE AND EXCHANGE RATE UNCERTAINTIES IN MINE DESIGN
where:
P is the price
α is the expected trend
σ is the standard deviation
dz is an increment in a standard Weiner process
dt is an increment of time
In Equation 2, µ is the logarithm of the long-term equilibrium
level of metal price and κ is the reversion speed. It is worth
noting here that the stochastic models in Equations 1 and 2 are
provided as examples for illustration purposes only. Other
models can replace the above ones without having effect on the
ideas and concepts presented in this work. The advantage of the
simple models presented in Equations 1 and 2 over the more
complex models is that the model parameters can be easily
estimated from historical data. More details about stochastic
models and parameter estimation can be found in Dixit and
Pindyck (1994) and Schwartz (1997), among others.
Valuation of mine designs
The economic valuation model is based on real options
valuations using the least squares Monte Carlo method. This
method was developed by Longstaff and Schwartz (2001) for
valuing American-style securities and extended by Abdel Sabour
and Poulin (2006) to valuing capital investments under multiple
market uncertainties and extended further by Dimitrakopoulos
and Abdel Sabour (2007) to valuing mining investments under
multiple market and geological uncertainties. In this work, the
operating flexibility to modify pit limits will be integrated into
the value estimate of designs taking into account metal prices,
exchange rates and geological uncertainties simultaneously.
As described above, metal prices and exchange rates
uncertainties can be quantified using the stochastic models
presented in Equations 1 and 2 while the geological uncertainty
is quantified using the conditional simulation technique. It is
assumed that there is M different feasible mine designs and each
design has H different grades and tonnages equal to the number
of generated orebody realisations. For market variables, a large
number of correlated realisations, N, are generated using the
stochastic models and taking into account the correlation
coefficients between those variables. Due to practical and
computational considerations the number of market realisations,
N, is much larger than the number of orebody realisations, H.
Therefore, to have N possible grades and tonnages for each
design throughout the mine life, N independent draws are
generated randomly from the H simulated orebody models. The
decision whether to keep or revise the originally defined pit
limits is taken at discrete regularly-spaced time points based on
the expected continuation value, CV, such as:
Keep the predefined pit limits if:
E (CV n, t) > 0 (3)
Revise the predefined pit limits if:
E (CV n, t) ≤ 0 (4)
According to Equations 3 and 4, if the expected present value
at time t and sample path n is positive, the optimum decision is to
keep the predefined pit limits until next decision time. Otherwise,
if the expected present value is negative, the original plan should
be revised and the current pit limit at time t should be the final
pit limit. Estimating the expected CV requires knowing the
function that relates the present value of future mining operations
beyond time t to the states prevailing at time t. This function
Advances in Orebody Modelling and Strategic Mine Planning I
could be in different forms such as the simple power series,
Laguerre polynomials or linear combinations of different forms
(Longstaff and Schwartz, 2001). The parameters of this function
can be estimated at each time by least squares regression.
This process is performed throughout the simulated paths N at
each period and the cash flows are defined based on the optimal
decision reached. If the time step is set to be one year, the
outputs of the valuation model are: the annual discounted cash
flows, the probability that the mine will be producing every year,
the expected value of positive annual cash flows with their
corresponding probabilities and the expected value of negative
annual cash flows with their corresponding probabilities. These
statistics are produced for each design in addition to the expected
overall present value for each design along with its confidence
limits at a specified level of confidence. These statistics are used
to rank the different designs as explained in the following
section.
Ranking of mine designs
The ranking procedure proposed in this study aims to gather
multiple value and risk analysis indicators into one quantitative
measure while integrating real industry complexities such as
uncertainty and operating flexibility to revise predefined pit
limits. The proposed design ranking system for selecting the best
open pit design based on the information available at the initial
planning time takes into account the following aspects:
• upside potential that measures the ability of designs to
capture possibly more profits than those expected if
outcomes were favourable;
• downside risk that reflects the difference between designs in
minimising negative cash flows risk throughout mine life;
• probability of completion, which is the probability that the
mine will be open throughout its planned life; and
• statistics of the estimated values which includes the average,
lower and upper limit at a certain confidence level.
After estimating the above described four indicators for each
mine design, the total ranking indicator is simply the summation
of these four indicators. The designs are then ranked according to
the total indicator and the design with the highest indicator
should be selected.
In this work, the average performance of all feasible mine
designs will be used to compare and rank these designs.
However, other references can replace these averages depending
the goals and objective specific to each mining project or a
mining company. Also, in this article, the above mentioned four
ranking criteria are given equal weights. In other applications,
different weights can be given to those criteria based on the
specific policy of a mining company.
CASE STUDY – SELECTING A DESIGN
FOR A CANADIAN COPPER MINE
In this section the proposed design ranking procedure will be
applied to a case study of a copper mine in Canada. It is assumed
that there are alternative technically feasible mine designs
available to decision-makers. These different mine designs could
be generated by a number of ways. One way, which is applied in
this study, is to use the method described in Dimitrakopoulos,
Martinez and Ramazan (2007) for generating different designs
from the multiple orebody simulations. These designs have to be
ranked so as to select the best one based on the information
available at the planning time. To investigate the efficiencies of
the valuation techniques and the proposed ranking system in
ranking alternatives under uncertainty, the mine designs will be
ranked based on the following four ranking measures:
• the expected value estimated by the conventional NPV
valuation method;
Spectrum Series Volume 17 31
S A ABDEL SABOUR and R DIMITRAKOPOULOS

• the NPV-based indicator explained above, except the TABLE 1

probability of completion indicator since the NPV of the Economic parameters for the copper mine.
mine calculated at the planning time before starting
production does not consider the flexibility to revise pit Item Description
limits in the future; Risk-free interest rate, % 9.20
• the expected value estimated by the real options valuation; Inflation, % 6.70
and Income taxes, % 40.00
• the real options valuation (ROV) -based indicator. Initial copper price, $US/lb 1.07
It is assumed that the mine was planned and mined out in the Volatility, %, copper price 20.00
past. Therefore, it is possible to compare the design ranking
based on the actual market realisations to the ranking estimated Reversion speed of copper price 0.19
by the four methods listed above. No capital expenditures are Long-term copper price, $US/lb 1.00
considered since it is assumed that production operations are Initial $US/$CAN rate 1.20
carried out by contractors.
Volatility, %, $US/$CAN 4.00
This copper mine is assumed to be planned in 1992, started
production at the beginning of 1993 and closed at the end of Reversion speed/year, $US/$CAN 0.08
2000. The economic data prevailing at the planning time (1992) Long-term, $US/$CAN 1.25
are listed in Table 1. Both of the copper price and the exchange
rate are modelled with the mean reverting process in Equation 2.
Except the initial copper price and the initial exchange rate, all with no loss offset. At the planning time, there were ten mine
parameter estimations are based the historical data from 1970 to designs estimated by feeding the ten simulated orebody models
1992. The overall income tax rate is assumed to be 40 per cent into Whittle Software that uses the nested Lerchs-Grossman

TABLE 2
Average annual tonnage, operating cost and grade for the technically feasible copper mine designs.

Y1 Y2 Y3 Y4 Y5 Y6 Y7 Y8
Design 1 Tonnage, tonne 7.40 7.50 5.93 7.50 7.50 7.46 7.50 4.89
Cost, $CAN/tonne 14.62 14.72 15.52 13.85 14.53 14.88 13.38 12.42
Copper grade, % 0.70 0.67 0.57 0.59 0.68 0.72 0.53 0.65
Design 2 Tonnage, tonne 7.50 7.50 7.35 7.50 7.50 6.21 7.50 3.75
Cost, $CAN/tonne 14.70 14.65 14.62 13.96 14.84 15.32 12.83 12.42
Copper grade, % 0.70 0.65 0.61 0.61 0.71 0.58 0.58 0.70
Design 3 Tonnage, tonne 7.50 7.50 7.47 6.45 7.50 7.08 7.07 5.45
Cost, $CAN/tonne 14.42 14.64 14.62 15.03 13.26 15.28 14.29 12.44
Copper grade, % 0.70 0.66 0.63 0.56 0.63 0.78 0.51 0.63
Design 4 Tonnage, tonne 7.50 7.50 7.47 5.04 7.50 7.50 7.29 5.89
Cost, $CAN/tonne 14.33 14.57 14.72 16.10 13.81 13.12 14.83 12.59
Copper grade, % 0.68 0.66 0.67 0.51 0.58 0.68 0.71 0.57
Design 5 Tonnage, tonne 7.50 6.99 7.48 7.50 7.50 5.99 7.50 4.86
Cost, $CAN/tonne 14.95 14.85 14.30 13.31 14.76 15.66 13.36 12.54
Copper grade, % 0.74 0.62 0.55 0.63 0.69 0.65 0.56 0.70
Design 6 Tonnage, tonne 7.50 7.50 6.09 7.50 7.50 7.33 7.50 2.95
Cost, $CAN/tonne 14.12 14.73 15.34 13.66 14.22 14.85 13.09 12.50
Copper grade, % 0.70 0.67 0.54 0.60 0.68 0.68 0.58 0.70
Design 7 Tonnage, tonne 7.50 7.50 7.50 6.00 7.50 7.26 7.50 4.07
Cost, $CAN/tonne 14.42 13.51 14.63 15.55 13.68 15.08 13.93 12.43
Copper grade, % 0.67 0.65 0.67 0.59 0.61 0.77 0.52 0.65
Design 8 Tonnage, tonne 7.50 6.83 5.20 7.50 7.50 7.50 7.50 6.13
Cost, $CAN/tonne 14.90 15.09 15.88 14.01 12.93 14.29 14.57 12.46
Copper grade, % 0.73 0.67 0.50 0.58 0.64 0.72 0.62 0.62
Design 9 Tonnage, tonne 7.50 7.50 6.92 7.50 7.50 6.74 7.41 4.54
Cost, $CAN/tonne 14.76 14.65 14.89 14.18 12.76 15.48 14.29 12.47
Copper grade, % 0.69 0.64 0.62 0.56 0.67 0.78 0.52 0.65
Design 10 Tonnage, tonne 7.46 7.50 7.50 6.75 7.50 7.44 7.50 3.07
Cost, $CAN/tonne 14.71 13.58 14.60 15.00 13.89 14.58 14.12 12.40
Copper grade, % 0.68 0.65 0.67 0.61 0.60 0.75 0.53 0.65

32 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 ACCOUNTING FOR JOINT ORE SUPPLY, METAL PRICE AND EXCHANGE RATE UNCERTAINTIES IN MINE DESIGN

algorithm to generate the optimum mine design (Whittle, 1999). 120

Table 2 lists the average annual tonnage, operating cost, in real
terms, and copper grades of the multiple simulated orebody 100
ROV
models for the ten designs. 80
Conventional NPV

Total ranking indicator

Each of the ten designs has been evaluated using Monte Carlo 60
method with 20 000 correlated simulations for the copper prices
and the exchange rates throughout the eight-year mine life. To 40
integrate geological uncertainty, 20 000 random independent 20
draws have been drawn from the ten simulated orebody
0
models. Therefore, there are 20 000 different possible scenarios
1 2 3 4 5 6 7 8 9 10
throughout the mine life with each scenario has its own copper -20
prices and exchange rates that are simulated using the stochastic
-40
models and an independent draw from the ten simulated ore
tonnages, production costs and ore grades. Two valuation -60
techniques are applied. The first technique is the conventional -80
NPV method that does not take into account operating flexibility Design
to revise pit limits. The second one is the ROV in which the
flexibility to revise pit limits is integrated and the optimum
decision is made based on the expected value conditional on time FIG 6 - Design indicators based on the NPV and the real
and the simulated paths. To focus on the value of operating options results.
flexibility only, the same discounting procedure was applied for
both the conventional NPV and the ROV methods. Therefore, the
difference between the two techniques here is owed solely to the TABLE 3
difference in the way they handle flexibility in pit design. Designs ranking for the copper mine.
Figure 5 shows the expected design values estimated by the
conventional static NPV and the ROV. Based on the NPV, Rank NPV NPV-based ROV ROV-based Actual
Design 6 has the highest value while the ROV indicates that expected TRI expected TRI data
Design 10 has the highest value. value value ranking
1 Design 6 Design 6 Design 10 Design 10 Design 2
45
ROV 2 Design 10 Design 10 Design 7 Design 2 Design 10
40 Conventional NPV
3 Design 7 Design 7 Design 2 Design 7 Design 7
35 4 Design 5 Design 5 Design 6 Design 3 Design 4
Value, $CAN million

30 5 Design 2 Design 2 Design 3 Design 4 Design 5

25 6 Design 1 Design 1 Design 5 Design 5 Design 3
7 Design 9 Design 9 Design 1 Design 1 Design 6
20
8 Design 3 Design 3 Design 4 Design 6 Design 9
15
9 Design 4 Design 8 Design 9 Design 9 Design 1
10
10 Design 8 Design 4 Design 8 Design 8 Design 8
5

0 while it is ranked the second based on the actual market data.

1 2 3 4 5 6 7 8
Design
FIG 5 - Designs valuation with the NPV and the real options.
The advantage of Monte Carlo method is that, it provides
multiple statistics for the overall value as well as annual cash
flows that can be used to carry out advanced risk analysis. These
statistics were used to construct a ranking system. Following the
proposed design ranking procedure outlined above, a NPV-based
indicator and a ROV-based indicator have been calculated for
each design. Figure 6 depicts the calculated Total Ranking
Indicators (TRI) based on the valuation results of NPV and ROV.
It is worth noting that the TRI for each design is the summation
of the subindicators related to the upside potential, downside
risk, probability of completion and value statistics. The only
difference between the NPV and the ROA indicators is that the
NPV one consists of three subindicators since the probability of
completion indicator is the same for all designs.
With the valuation results in Figure 5 and the indicators in
Figure 6, it is possible to rank the ten mine designs based on the
four different measures, as shown in Table 3. The last column lists
design ranking based on the actual market data over the mine life
(Y 1992 - Y 2000). As indicated in Table 3, Design 6 is ranked the
first based on NPV, while based on the actual market data it is
ranked the seventh. Based on ROV, Design 10 is ranked the first
9 10
Comparing the rank for all other designs, it is obvious that design
ranking based on ROV is closer to the ranking based on actual
market data. Another important conclusion can be drawn from
Table 3 is that, for the conventional NPV analysis, design ranking
based on the NPV expected value is almost identical to that of the
NPV-based indicator and differs from those of the real options
expected value and the ROV-based indicator. The ROV expected
value and the ROV-based indicator have similar ranking for five
mine designs and different ranking for the remaining five designs.
The results show that under the conditions of uncertainty, design
values based on actual market data can be significantly different
from those estimated at the planning time. Consequently, using the
expected value to rank possible mine designs may result is
suboptimal decisions. Therefore, it is important to integrate
multiple risk and cash flow analysis into mine design selection
process. However, given the multiple sources of uncertainty
associated with mining investments, the multiple risk and cash
flow analysis could be useless if the management flexibility to
react to the new information is not considered. As shown in
Table 3, the NPV-based indicator using multiple risk and cash flow
analysis generated almost the same decision as the NPV expected
value. This is because if no flexibility to revise pit limits is
considered, the distribution of any estimated measure may
assumed to be approximately symmetrical around the expected
value and no significant change in the decision nor any
improvement in the selection process might be achieved. On the
contrary, when integrating one kind of operating flexibility, which

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 33
S A ABDEL SABOUR and R DIMITRAKOPOULOS
is the flexibility to revise the originally taken decisions regarding
the ultimate pit limits, the efficiency of the selection process was
significantly improved when considering multiple risk and cash
flow analysis measures. Effort has to be devoted to integrate more
risk analysis measures as well as more operating flexibilities, such
as the flexibilities to expend, contract, switch cut-off grades and
switch production inputs in order to improve further the process of
mine planning under uncertainty.
The results of this study are specific to the presented case
study and the data fed into the valuation model. Different results
might be obtained if the mine was assumed to be started and
finished at different calendar times. Also, the valuation results
are sensitive to the chosen stochastic model for describing the
evolution of a market variable and its estimated parameters. In
this respect, it is worth stressing that the ROV is a valuation
rather than a price forecasting technique. Like the static NPV, or
even any other system, the outputs are dependent on the inputs.
CONCLUSIONS
A ranking system was outlined for selecting open pit mine designs
using multiple risk and cash flow analysis measures while
integrating market and geological uncertainty and the operating
flexibility to revise pit limits. The proposed system is based on
valuing the possible, technically feasible, mine designs using a
simulation-based real options valuation method that can integrate
multiple sources of uncertainty related to market and geological
variables as well as the management responses to the new
information in the future. This economic valuation model
generates multiple statistics for cash flows and operating modes of
each design throughout the active life of the project. Such statistics
are used to construct a system for design ranking under
uncertainty.
For the sake of comparison, the proposed system was applied
along with other three ranking methods, based on the NPV
expected value, the ROV expected value and the NPV-based
indicator, to rank possible designs at a copper mine. The results
show that the design selection process may improve when
incorporating uncertainty and operating flexibility. Future
extensions could include integrating more risk analysis measures
and other types of management flexibilities such as the flexibility
to revise the cut-off grade with time so as to improve the design
selection process even further.
ACKNOWLEDGEMENTS
The work in this paper was funded from NSERC CDR Grant
335696 and BHP Billiton, as well NSERC Discovery Grant
239019, McGill’s COSMO Lab and its industry members,
AngloGold Ashanti, Barrick, BHP Billiton, De Beers, Newmont,
Vale and Vale Inco.
REFERENCES
Abdel Sabour, S A and Poulin, R, 2006. Valuing real capital investments
using the least-squares Monte Carlo method, The Engineering
Economist, 51:141-160.
Brennan, M J and Schwartz, E S, 1985. Evaluating natural resource
investments, Journal of Business, 58:135-157.
Dimitrakopoulos, R, 2007. Applied risk assessment for ore reserves and
mine planning, Professional development short course notes, 350 p
(The Australasian Institute of Mining and Metallurgy: Melbourne).
34 Spectrum Series Volume 17
Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans
under uncertainty: Can the real options make a difference?,
Resources Policy, 32:116-125.
Dimitrakopoulos, R, Farrelly, C T and Godoy, M, 2002. Moving forward
from traditional optimisation: Grade uncertainty and risk effects in
open pit design, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 111:A82-A88.
Dimitrakopoulos, R, Martinez, L S and Ramazan S, 2007. A maximum
upside/minimum downside approach to the traditional optimisation
of open pit mine design, Journal of Mining Science, 43:73-82.
Dixit, A K and Pindyck, R S, 1994. Investment Under Uncertainty
(Princeton University Press: New Jersey).
Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling, SME Transactions,
316:43-50.
Kamrad, B and Ernst, R, 2001. An economic model for evaluating mining
and manufacturing ventures with output yield uncertainty, Oper Res,
49(5):690-699.
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model
for open pit mine planning: Application and risk analysis at a copper
deposit, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 116:A109-A118.
Longstaff, F A and Schwartz, E S, 2001. Valuing American options by
simulation: A simple least-squares approach, The Review of
Financial Studies, 14:113-147.
Moel, A and Tufano, P, 2002. When are real options exercised? An
empirical study of mine closings, Review Financial Studies,
15(1):35-64.
Monkhouse, P H L and Yeates, G, 2007. Beyond naive optimisation, in
Orebody Modelling and Strategic Mine Planning, second edition (ed:
R Dimitrakopoulos), pp 3-8 (The Australasian Institute of Mining and
Metallurgy: Melbourne).
Ramazan, S, 2007. The new fundamental tree algorithm for production
scheduling of open pit mines, European Journal of Operations
Research, 177 1153-1166.
Samis, M, Davis, G A, Laughton, D and Poulin, R, 2006. Valuing
uncertain asset cash flows when there are no options: A real options
approach, Resources Policy, 30:285-298.
Samis, M, Laughton, D and Poulin, R, 2003. Risk discounting: The
fundamental difference between the real option and discounted cash
flow project valuation methods [online]. Available from: <http://
papers.ssrn.com> [Accessed: 18 November, 2008].
Samis, M and Poulin, R, 1998. Valuing management flexibility: A basis
to compare the standard DCF and MAP frameworks, CIM Bulletin,
91:69-74.
Schwartz, E S, 1997. The stochastic behaviour of commodity prices:
Implications for valuation and hedging, Journal of Finance,
52:923-973.
Slade, M E, 2001. Valuing managerial flexibility: an application of real
option theory to mining investments, Journal of Environmental
Economics and Management, 41:193-233.
Tufano, P and Moel, A, 1999. Bidding for Antamina: Incentives in a real
option context, in Project Flexibility, Agency and Competition, (eds:
M J Brennan and L Trigeorgis), pp 128-150 (Oxford University
Press: Oxford).
Vallee, M, 2000. Mineral resource + engineering, economic and legal
feasibility = ore reserve, CIM Bulletin, 93:53-61.
Whittle, J, 1988. Beyond optimisation in open pit design, in Proceedings
Canadian Conference on Computer Applications in the Mineral
Industries, pp 331-337 (Balkema: Rotterdam).
Whittle, J, 1999. A decade of open pit mine planning and optimisation –
The craft of turning algorithms into packages, in Proceedings
APCOM ’99 – Computer Applications in the Minerals Industries
28th International Symposium, pp 15-24 (Colorado School of Mines:
Golden).
Advances in Orebody Modelling and Strategic Mine Planning I

Production Scheduling Under Metal Uncertainty – Application of
Stochastic Mathematical Programming at an Open Pit Copper
Mine and Comparison to Conventional Scheduling
A Leite1 and R Dimitrakopoulos2
ABSTRACT
Life-of-Mine (LOM) production scheduling of open pit mines is of major
importance as it defines the ore supply that the mine can provide while
maximising its present value. The conventional mine scheduling approach
requires a single, average type representation of reality and, as a result, it
ignores uncertainty about technical and economic parameters. From all
sources of uncertainty affecting the planned ore supply of a given
schedule, uncertainty about ore reserves is acknowledged to be the most
critical one. It is known that the use of a single, average type of deposit
model to define reserves produces a smoothed image of the real deposit
and this smoothing may lead to a misleading forecast of ore supply, waste
and related cash flow over the life of the mine. To overcome this problem,
a stochastic mining production scheduling approach is considered here.
The approach explicitly integrates the uncertainty of parameters or inputs
to the scheduling problem aimed at generating a more risk-robust
solution. In this study, a Stochastic Integer Programming (SIP)
formulation for mining production scheduling is applied and tested at a
low-grade variability copper deposit. The stochastic solution as
implemented herein aims to maximise the economic value of a project
and minimise deviations from production targets in the presence of
geological uncertainty. Unlike the conventional approach, the method is
able to account for and manage risk.
In the case study presented herein, the stochastic formulation for mine
production scheduling is shown to produce 29 per cent higher Net Present
Value (NPV) than the schedule obtained from a conventional scheduler.
This is an important difference, particularly under the current global
financial climate.
INTRODUCTION
The optimum open pit mine production schedule is defined as the
sequence of extraction that maximises the present value of the
project. This task is one of the most challenging and important in
the mine planning framework as it defines the ore supply
produced over the LOM and consequentially has a substantial
impact on the NPV of the project. The conventional mine design
and production framework defines the extraction sequence
considering a single, average type of orebody model as input. As
a result, it does not account for uncertainty in the related project
parameters and leads to misleading assessments (Ravenscroft,
1992; Dimitrakopoulos, Farrelly and Godoy, 2002). This finding
makes clear that there is a need to address the mine production
schedule problem using stochastic approaches, which use as
input a set of equally probable representations of the orebody
reproducing, by construction, its actual spatial variability and
distribution. This set of representations is able to directly
incorporate uncertainty into the formulation of the problem.
Different stochastic approaches have been considered in order
to provide more robust solutions dealing with uncertainty.
Dimitrakopoulos, Martinez and Ramazan (2007) utilise the
concepts of upside/downside potential to include grade uncertainty
in pushback design. Several pit designs for a set of simulated
orebodies are obtained and a final pushback configuration selected
1. Vale Ltd, Canada. Email: Andre.Leite@vale.com
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montréal QC H3A 2A7, Canada.
Email: roussos.dimitrakopoulos@mcgill.ca
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
by considering the one with a minimum downside or maximum
upside potential. Dimitrakopoulos and Ramazan (2004) present a
mathematical programming formulation minimising deviations
form production targets by considering a probabilistic approach.
The new concept of geological discounted rate is introduced and
produces a decreasing unit cost for deviation of target over the
LOM. Godoy and Dimitrakopoulos (2004) and Leite and
Dimitrakopoulos (2007) propose a different stochastic approach
using a simulated annealing algorithm to obtain a stochastic
schedule. The proposed solution is divided into three stages:
1. optimum mining rates are defined using an LP formulation,
2. rates are utilised to schedule a set of simulated orebodies,
and
3. schedules are then used in a final stage in which a
stochastic schedule is obtained by using a simulated
annealing algorithm and the schedules as an input.
The studies show a 25 per cent to 28 per cent difference in
NPV as compared to the conventional schedule. Menabde et al
(2007) implement a stochastic integer programming (SIP)
formulation for long-term production schedule that maximises
NPV considering several possible simulated orebodies and
simultaneously optimising cut-off grades. Ramazan and
Dimitrakopoulos (2007a) propose an approach that accounts for
all available realisations of the orebody simultaneously in an SIP
formulation. Their formulation has as objectives the maxim-
isation of NPV and minimisation of deviation from production
targets. Different penalties may be defined for deviations of
different targets. Jewbali (2006) combines the use of the SIP
formulation with the use of simulated future data, updating
simulated models to produce an optimum stochastic mine
schedule and shows a 30 per cent increase in NPV at a gold
deposit.
Despite substantial monetary benefits, the application of
stochastic schedulers is relatively recent and the value of this
solution when applied to different types of deposits is still not
completely understood. In the present study, the approach
presented by Ramazan and Dimitrakopoulos (2007b) is applied
to a low-grade variability copper deposit. The study tests the
approach, quantifies the associated value of the stochastic
solution, assesses the risk profile of pertinent mining parameters
and, finally, analyses the results to propose future improvements.
The following sections outline the stochastic integer
programming formulation, present its application in a copper
deposit, and compare the results with those obtained by a
conventional scheduler. Conclusions follow.
A Stochastic Integer Programming formulation
Stochastic mathematical programming approaches to the mine
schedule problem are avenues to directly incorporate uncertainty
about ore supply in the formulation of the problem, so as to
minimise risk of not meeting the mine production targets. In the
mine production schedule case, the decision to be made is the
time period in which each block is mined, in order to maximise
the overall discounted value of the project, subject to slope,
reserves and processing and mining capacity constraints. The set
Spectrum Series Volume 17 35
A LEITE and R DIMITRAKOPOULOS

of blocks available to be scheduled are the ones contained within production target deviations. It can also be used to blend risk by
the ultimate pit. The SIP formulation presented herein includes using a low geologic risk discount rate.
uncertainty in the formulation of the problem by considering a The unit costs in Equation 1 are for excess production:
set of different and equally probable stochastic simulated
orebody realisations in the optimisation process. cu0 o
cuto = (2)
(1 + GRD)t
Objective function
The SIP formulation below aims to maximise the NPV of the where:
mine by minimising the risk of falling short of previously defined d is the discount factor mentioned earlier, or for deficient
production targets. It includes two possible destinations for a production
block:
1. processing plant, or cl0 o
clto = (3)
2. waste dump. (1 + GRD)t
The objective function includes two components and it is:
Equation 1 is subject to constraints detailed in Ramazan and
Dimitrakopoulos (2007b).
⎡ ⎤
p ⎢N m ⎥
Max ∑ ⎢∑ E {( NPV )i} * bi − ∑ ( cu d su + cl d sl ⎥ APPLICATION AT A COPPER DEPOSIT
t t to to to to
(1)
⎢1 144244 3⎥
t=1 i=l
442443 s=1
Part B
⎢⎣ Part A ⎥⎦ The deposit
The deposit is located in a typical archean greenstone belt. The
where:
region consists predominantly of mafic lavas with lesser amounts
i is the block identifier of intermediate to felsics volcaniclastics. The geological database
t is the time period; to flags the ore production target type consists of 185 drill holes with 10 m copper composites in a
pseudo-regular grid of 50 m × 50 m covering an approximately
l stands for lower bound rectangular area of 1600 × 900 m2. Using the geological inform-
u stands for upper bound ation available, one mineralisation domain is defined and modeled
through a geostatistical study. An estimated model, obtained using
s stands for the simulation number ordinary kriging, and a set of simulated models, generated using a
p is the maximum number of scheduling periods direct block simulation method (Godoy, 2003) are produced. All
models are constrained by the ore domain previously defined.
N is the total number of blocks to be scheduled Figure 1 brings a section of the estimated model and three of the
bit is a variable representing the portion of block i to be mined simulated ones. The smoothing effect on the grade, produced by
in period t the ordinary kriging model, is made clear in Figure 1.
If defined as a binary variable, it is equal to one if the block i
is to be mined in period t and equal to 0. Otherwise;
E {( NPV )ti} is the expected NPV to be generated by mining
block i in period t – it is computed as the
discounted value of Equation 1
cuto is the unit cost for excess of ore production
to
d su is the excess amount of ore production in period t
considering simulation s
clto is the unit cost for the deficient ore production
to
d sl is the deficient amount of ore production in
period t considering simulation s
The first component (Part A) in Equation 1 contributes to the
maximisation of NPV of the project. The expected NPV of a block
is computed as the expected present value if the block is mined in
period t, considering all simulated values. The second part (Part B) FIG 1 - East-West sections of three simulated models (left column)
is responsible for minimising deviations from ore production and the corresponding single estimated model (right).
targets, also managing the distribution of risk within and between
periods over the LOM. Risk management is accomplished by the
use of a geological discount rate (GDR), which discounts over Production scheduling
time the penalties applied to the unit cost deviations as explained The SIP formulation in Equation 1 used to generate the mine
below. The initial penalties for excess, cu0 o , or shortage production, production schedule first requires the definition of the ultimate
cl0 o , are user-defined and should be at the same order of magnitude pit and pushbacks. The nested pit implementation of the
as the first part of the objective function to ensure the second part Lerchs-Grossman algorithm (Whittle, 1988) for pit optimisation
is being properly considered. The impact of discounted penalties is is used here to define a final pit and pushbacks. This algorithm
a progressive decrease in the unit cost over the periods. This setup requires as an input a single representation of the orebody, which
can be used to ensure that less ‘risky’ mining blocks will be is created using ordinary kriging on a 20 × 20 × 10 m3 block
scheduled in early periods, therefore decreasing the risk of not support. The associated economic and technical parameters are
attaining the planned targets and guarantying the minimisation of given in Table 1. This model and parameters are utilised in the

36 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 PRODUCTION SCHEDULING UNDER METAL UNCERTAINTY

following sections for comparison with the results obtained by additional schedules are generated, as one of the objectives of the
the SIP. work is to test and better understand the effect of the geological
Using the parameters in Table 1 and the conventionally discount rate on the schedule produced.
estimated orebody model, a set of nested pits is generated. Pit 16
is selected as the ultimate pit limit as it corresponds to the
maximum net present value. There are 14 480 blocks inside the pit
limits. In this study, a previously established ore processing
capacity of 7.5 M tonnes per year is used. The yearly maximum
mining capacity is set to 28 M tonnes although there is no
constraint to guarantee a constant material movement over the
LOM. With this setup the scheduler is free to define the optimum
waste production strategy. Twenty realisations of the deposit are
available and generated with the direct block simulation method
(Godoy, 2003). To obtain a simulated model, the orebody is
divided into blocks of 20 × 20 × 10 m3 within the mineralised
domain. Each block is then represented by 10 × 10 × 1 nodes. This
number of nodes, 100 per block, is large enough to ensure that the
actual block scale variability is reproduced by the simulated
orebodies. These stochastic simulated orebodies are used as an
input for the SIP models and to produce risk profiles of
performance parameters of the schedules generate throughout the
study.
FIG 2 - Risk profile for ore production of the base case schedule.
TABLE 1
Economic and technical parameters.
Copper price (US$/lb) 2.0
Selling cost (US$/lb) 0.3
Mining cost ($/tonne) 1.0
Processing cost ($/tonne) 9.0
Slope angle 45°
Processing recovery 0.9

A base case SIP schedule is first considered and then the

sensitivity of pertinent key project parameters is tested and their
impact on the project’s NPV and performance evaluated. The
stochastic schedule is generated considering a 20 per cent
geological discount rate. All cases utilise the same parameters as
specified in Table 2. A fixed copper grade cut-off of 0.3 per cent is
used in combination with the probability cut-off to classify ore and
waste blocks. Penalties for excess and shortage production are
select accordingly to the magnitude of the first part of the
objective function in order to ensure that the second component, FIG 3 - Risk profile for waste production of the base case
which accounts for minimisation of production targets and risk schedule.
management, is properly weighted in the objective function. Past
work has shown that it is the order of magnitude of the penalties
rather than the actual values that effect the optimisation process. A
higher penalty for shortage production is imposed as it has a more
severe impact on the project than an excess production, which
could be handled by the use of stockpiles. Risk profiles for ore and
waste production, cumulative NPV and production deviations are
generated and presented respectively by Figures 2, 3, 4 and 5. Two

TABLE 2
SIP related parameters.

Geologic risk discount rate 20%

Cost of shortage in ore production 10 000
(unit/tonne)
Cost of excess ore production (unit/tonne) 1000
Economic discount rate (%) 10
Cut-off (% Cu) 0.3
Number of simulated orebody models 20 FIG 4 - Risk profile for cumulative NPV of the base case schedule.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 37
A LEITE and R DIMITRAKOPOULOS

related risk profile that the conventional schedule generated

misleading results in the presence of grade uncertainty – the
forecasted ore supply has about a five per cent chance (one in 20
chances that equally likely scenarios of the deposit will produce
what is expected) for almost all years. Only in two years is the
expected forecasted production realised, year 5 and year 7. The
economic implication of such deviations can be seen in Figure 7,
where the risk profile of cumulative NPV for both the base case
stochastic schedule and the conventional schedule are shown.
The average NPV difference between the conventional and the
stochastic solution is approximately 29 per cent. The higher NPV
obtained by the stochastic solution is obtained by first
incorporating grade uncertainty into the mine production
scheduling formulation, minimising the possible deviations of ore
production target and at the same time managing the risk between
mine periods. The difference between the value of the stochastic
solution and the conventional one comes from the capacity of the
stochastic optimiser to obtain an optimum schedule considering
simultaneously several equally probable orebody models. The
FIG 5 - Deviations from ore production targets for the base case equally probable orebody models, obtained by conditional
schedule. simulation (better than the smoothed model utilised by the
conventional scheduler), represent the spatial grade distribution
of the deposit.
COMPARISON TO A CONVENTIONAL
SCHEDULER
To assess the value of the stochastic solution in the case study
presented above, a conventional LOM production schedule is
developed in this section and the economic difference between
the stochastic and conventional approaches are evaluated. The
conventional schedule is generated using the average type of
deposit model previously described and obtained using the
technique of ordinary kriging. The same final pit and push-backs
are utilised with the associated technical and economic
parameters as before (Table 1). The schedule is obtained using
Millawa NPV algorithm (Whittle, 1999) using a fixed 7.5 M
tonnes ore production target and a maximum material movement
of 28 M tonnes. No constraint or limitation is imposed on the
scheduler, so as to keep the total material movement constant
over time. The intent is to let the optimiser decide the best waste
removal strategy and to replicate the condition imposed on the
SIP scheduler.
Figure 6 presents the ore production of the conventional FIG 6 - Ore production and associated risk profile for the
schedule and its risk profile. It is clear in the figure and the conventional schedule.

FIG 7 - Conventional and SIP base case schedules and their respective risk profiles.

38 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I

The conventional scheduler forecasts a LOM of eight years in
contrast with the seven years LOM forecasted by the stochastic
scheduler. By using a smoothed scenario, the conventional
scheduler overestimates the amount of ore above the 0.3 % Cu
cut-off, therefore the one extra year of production is required to
fully mine the deposit. It is interesting to emphasise that this fact
cannot be generalised as it is a function of the grade distribution
and cut-off applied.
CONCLUSIONS
This study shows a stochastic mine scheduling formulation and its
application to scheduling under grade uncertainty. The results find
the value of the stochastic solution by comparing the results
obtained by the stochastic schedule with the one obtained by a
conventional scheduler. The difference is significant (29 per cent)
and derived from the inability of the conventional scheduler to
include grade uncertainty in its formulation, in comparison with
the stochastic scheduler which not only capable of accounting for
the grade uncertainty in its formulation but also of managing the
risk associated with it in the deviation of production targets.
The fact that a stochastic integer formulation solution of the
mine production schedule problem generates a higher value
reflects the importance of incorporating geological uncertainty
into the scheduler problem formulation. Similar findings are
presented by Leite and Dimitrakopoulos (2007), which tested a
different stochastic scheduler using the same copper deposit.
Like other available past studies, both studies make clear the
advantages of using a stochastic approach in the mine production
schedule formulation.
ACKNOWLEDGEMENTS
The work in this paper was funded from NSERC CDR Grant
335696 and BHP Billiton, as well NSERC Discovery Grant
239019. Thanks are in order to Brian Baird, Peter Stone and
Gavin Yates, as well as BHP Billiton Diamonds and, in
particular, Darren Dyck, for their support, collaboration and
technical comments.
REFERENCES
Dimitrakopoulos, R, Farrelly, C T and Gody, M, 2002. Moving forward
from traditional optimisation: Grade uncertainty and risk effects in
open-pit design, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 111:A82-A88.
Advances in Orebody Modelling and Strategic Mine Planning I
PRODUCTION SCHEDULING UNDER METAL UNCERTAINTY
Dimitrakopoulos, R, Martinez, L S and Ramazan, S, 2007. A maximum
upside/minimum downside approach to the traditional optimisation
of open pit mine design, Journal of Mining Science, 43:73-82.
Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty based production
scheduling in open pit mining, SME Transactions, 316:106-112.
Godoy, M, 2003. The efficient management of geological risk in
long-term production scheduling of open pit mines, PhD thesis,
University of Queensland, Brisbane.
Godoy, M and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling, SME Transactions,
316:43-50.
Jewbali, A, 2006. Modelling geological uncertainty for stochastic
short-term production scheduling in open pit mines, PhD thesis,
University of Queensland, Brisbane.
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model
for open pit mine planning: Application and risk analysis at a copper
deposit, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 116:A109-A118.
Menabde, M, Froyland, G, Stone, P and Yeates, G, 2007. Mining
schedule optimisation for conditionally simulated orebodies, in
Orebody Modelling and Strategic Mine Planning, second edition
(ed: R Dimitrakopoulos), pp 379-384 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Ramazan, S and Dimitrakopoulos, R, 2007a. Stochastic optimisation of
long-term production scheduling for open pit mines with a new
integer programming formulation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 385-392 (The Australasian Institute of Mining and Mettalurgy:
Melbourne).
Ramazan, S and Dimitrakopoulos, R, 2007b. Production scheduling with
uncertain supply: A new solution to the open pit mining problem,
Cosmo Research Report, volume 2, pp 257-294.
Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
simulation, in Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 101:A104-A108
Whittle, J, 1988. Beyond optimisation in open pit design, in Proceedings
Canadian Conference on Computer Applications in the Mineral
Industries, pp 331-337 (Balkema: Rotterdam).
Whittle, J, 1999. A decade of open pit mine planning and optimisation –
The craft of turning algorithms into packages, in Proceedings 28th
International Symposium on the Application of Computers and
Operations Research in the Mineral Industry, pp 15-24 (Colorado
School of Mines: Golden).
Spectrum Series Volume 17 39

Regional Three-Dimensional Modelling of Iron Ore Exploration
Targets
V Osterholt1, O Herod2 and H Arvidson3
ABSTRACT
The Australasian Code for Reporting of Mineral Resources and Ore
Reserves, December 2004 (the JORC Code) deals with reporting
exploration target size and type, which are expressed as ranges in order to
properly convey the high level of uncertainty that exists early in the
project evaluation process. To inform investors of the significant potential
mineralisation in BHP Billiton’s Pilbara tenements and to enable strategic
planning, a rigorous modelling approach has been devised for such early
stage projects.
An innovative approach to modelling scenarios that capture the
uncertainty range has been developed. The approach integrates drill holes,
mapping, geophysical surveys and the geologist’s interpretation. The
modelling workflow consists of two parts: geology modelling, and grade
modelling. An implicit 3D modelling approach is utilised to rapidly
generate multiple regional models of stratigraphy. Mineralisation envelopes
and weathering horizons, providing further controls on the target range, are
also modelled using fit-for-purpose tools and data structures. Grades are
estimated with a recoverable resource estimation approach. The estimation
accounts for uncertainty of local grade distributions and provides
multi-variate grade models that honour correlations between variables. The
final product delivers a rigorous approach to the estimation of tonnage and
grade ranges for reporting the exploration target that captures the
uncertainty of the estimate.
INTRODUCTION
In the Hamersley Ranges, bedded iron ores comprise the bulk of
the area’s iron enrichment, varying between low phosphorous (P)
martite-microplaty haematite and medium-high P martite-
goethite or martite-ochreus goethite ores, originating from
enriched banded iron-formation (BIF) sequences of the
Brockman and Marra Mamba Formations. Stratigraphy is folded,
thrusted and faulted to various degrees across the Hamersley
Ranges as a result of multiple deformation phases. Structural
features often control the quantity and quality of mineralisation
by influencing the available volumes of BIF as well as the
enrichment processes. Figure 1 shows an example geological
section through an openly folded orebody, with mineralisation
boundaries subparallel to stratigraphy and high continuity
throughout the syncline. Weathering effects often truncate
potentially economic mineralisation near the surface. Data
available during early evaluation work is usually sparse, with
only some drill holes per kilometre along strike, geological
surface mapping and geophysical data. Often data is historic and
associated with data entry errors, low precision, old drilling
techniques and can be incomplete. Such data sets are not
considered sufficient to support public reporting of mineral
resources, which is an expression of the inherent uncertainty. In
order to advance exploration and strategic planning, this
uncertainty around exploration targets needs to be addressed in a
1. MAusIMM, Senior Resource Geologist, BHP Billiton Iron Ore,
PO Box 7122, Cloisters Square, Perth WA 6850.
Email: Volker.Osterholt@bhpbilliton.com
2. MAusIMM, Senior Consultant, SRK Consulting, PO Box 943,
West Perth WA 6872. Email: OHerod@srk.com.au
3. MAusIMM, Principal Resource Geologist, Resource Modelling
Group, BHP Billiton Iron Ore, PO Box 7122, Cloisters Square,
Perth WA 6850. Email: Heath.Arvidson@bhpbilliton.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
meaningful and practical way. Estimates of such potential
mineralisation can be developed in a variety of ways.
Historically, BHP Billiton has generated these as a deterministic
non-spatial grade and tonnage estimate. In the presence of
uncertainty, deterministic estimates do not deliver the required
exploration target ranges. They also have limited use in strategic
planning, particularly in bulk commodities like iron ore, where
multi-pit blending options are a key strategic driver.
MODELLING APPROACH
A holistic view of geological uncertainty requires capturing the
uncertainty about the key factors controlling the grade-tonnage
distribution of a prospective area. A target evaluation project
consists of three stages:
1. development of geology and grade scenarios;
2. geology modelling to build 3D representations of stationary
domains, that is volume models; and
3. grade modelling to populate volumes with the relevant
grade estimates.
The initial conceptual analysis is based on understanding the
states of knowledge (facts, assumptions, interpretations and
unknowns), followed by describing the linkage between key
influences on mineralisation. Concepts are then distilled into a
range of possible scenarios, a best-guess scenario (most likely)
and alternative pessimistic and optimistic scenarios (less likely
but still reasonably likely as an outcome). These scenarios are
documented in a case matrix. This case matrix is the driving tool
of the modelling process – it links data analysis, interpretation
and scenarios developed during the conceptual analysis. All
relevant details for all key drivers for the model cases are listed
in a transparent and auditable way. A simple example illustrating
the principle is given in Figure 2.
Developing scenarios and assigning probabilities is an expert-
driven process including team workshops to utilise all available
sources of knowledge. Detailed data analysis and a review of
previous work and reports is combined with regional geological
knowledge and experience from existing mining operations, which
are then used to develop a range around the key influences.
Examples include mineralisation in particular units of the
stratigraphic sequence, occurrence of (un-) favourable structures,
continuity of the mineralisation along strike and down-dip and
weathering intensity. The necessity and challenge of such expert
judgements for early stage resource uncertainty assessment was
raised previously (eg Caumon and Caers, 2004). An innovative
approach to modelling these scenarios in three dimensions, with
the uncertainty, has been developed – it integrates drill holes,
mapping, geophysical surveys and the geologist’s interpretation.
This paper describes the modelling approach in a generic way.
Results presented in figures provide an illustration of typical study
outcomes. The workflow consists of two parts, geology modelling
and grade modelling.
These are presented in the next two sections, then the resulting
modelling capabilities are summarised, and finally, the strengths
of the approach and challenges are discussed.
Spectrum Series Volume 17 43
V OSTERHOLT, O HEROD and H ARVIDSON

FIG 1 - Geological section of brockman iron-formation hosted orebody (from unpublished internal BHP Billiton report).

FIG 2 - Influence diagram with key parameters (yellow) (LHS) and case matrix of the potential mineralisation of a region (RHS).

MODELLING OF GEOLOGY mineralisation are triangulated to construct 3D domain models for

The basis for spatial modelling of potential mineralisation is the
geology, which in most non-primary mineral systems can be
expressed as the underlying lithology, with some form of
mineralisation system superimposed. In data-rich (typically well
drilled) areas, sectional interpretations of the lithology and
use in grade estimation. Using traditional mine planning software,
the time to construct and subsequently update these models can be
significant and results in single geological models that are updated
annually for reporting purposes. When considering exploration
areas, the modelling time can often increase as additional non-data
constrained interpretations are required to produce appropriate

44 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 REGIONAL THREE-DIMENSIONAL MODELLING OF IRON ORE EXPLORATION TARGETS
triangulation models. To address the range of possible geological
interpretations at exploration stages, a different approach to
geological modelling is needed. Implicit modelling tools are
capable of rapidly building geologically robust triangulation
models of lithology and form the basis of a geological modelling
workflow presented in Figure 3. Modelling zones of potential
mineralisation forms the second part of the workflow, including
combining data (where available) with a method of representing
the understanding of the mineralisation processes.
Modelling of lithology
Projects at exploration stages typically have surface geological
mapping, comprising lithological contacts and orientation data and
some often shallow drilling. The project geologist will have a
conceptual understanding of both the spatial distribution of the
lithology and controls on the mineralisation system. The implicit
geological modelling tools used in this study combine a set of
geological rules (in the form of a stratigraphic sequence) with
geological interface and orientation data to construct a series of
3D functions which can be evaluated to create triangulations (see
Calcagno et al, 2008 for details of the methodology). The tools are
particularly appropriate for layered stratigraphy and thus very
applicable for modelling banded iron ore deposits.
The advantage of using implicit modelling tools is that they are
quick, with triangulations being generated within a few hours.
However, they do not represent a silver bullet to geological
modelling. Input and control by a geologist needs to be applied to
assure reasonable results. This typically takes the form of:
FIG 3 - Flow diagram showing a standard geological modelling workflow for resource estimation compared to a geological modelling
workflow using implicit modelling and separate mineralisation potential modelling to generate a series of triangulation models.
Advances in Orebody Modelling and Strategic Mine Planning I
• reviewing consistency between different data types,
• introducing additional interpretation data (ie limited control
sections), and
• simplifying the interpretation in data rich areas.
Through all the modelling, particular attention needs to be
placed on the scale of the data with respect to the scale of the final
model. It is not possible to precisely fit the triangulations to very
closely spaced data or detailed interpretations in structurally
complex areas without post-modelling manual adjustment. In all
cases where these modelling approaches have been applied, the
conceptual understanding of the geology has improved, as the
geologist can quickly explore the implications of different
interpretations. By applying different geological interpretations
derived by an expert panel (eg increased fold amplitude or reduced
throw on a fault), a series of lithological models can be produced
relatively quickly. These can then be used as the basis for
mineralisation modelling and subsequent grade estimation.
Modelling the mineralisation
Within the lithology models, it is necessary to identify zones of
(potential) mineralisation. In mapped areas, mineralisation may
have been identified at surface, and assay results from drilling
will clearly indicate subsurface mineralisation. However, away
from this data, a method was required to highlight areas of
potential mineralisation based on an understanding of the
mineralisation process. This is achieved through a combination
Spectrum Series Volume 17 45
V OSTERHOLT, O HEROD and H ARVIDSON
of suitable estimation of data in a 3D grid model and selecting
zones of the model based on modelled geological features. A
mineralisation indicator was applied to the drilling data
(continuous values) to combine it with mapping (mineralised,
un-mineralised or unknown, ie binary or null). The parameters of
the estimation of the combined data (distance, direction, number
of data required and method of estimation) was determined by an
expert panel and varied depending on the scenario. By using
deformable 3D grids controlled by the lithology model, the
estimation was able to honour the structural complexity of the
stratigraphy. Away from the data, potentially mineralised areas of
the grid model were selected by application of a conceptual
mineralisation model. A series of 3D geographical information
system (GIS) tools were used which could query regions of the
model, eg those regions within 100 m of faults or occurring in
synformal fold hinges. The specific mineralisation rules were
part of the expert panel scenario. The initial spatial visualisation
of these ‘rules’ often identified unrealistic assumptions, which
could then be re-evaluated by the geologists involved.
Combining the two approaches gives a grid model with a
series of mineralised zones – the zones can be triangulated for
visualisation purposes. The application of indicator estimation
and a series of mineralisation ‘rules’ provides a repeatable and
quick method of generating zones of potential mineralisation
which are controlled by the underlying lithology model. For each
lithology model, a series of mineralisation models could be built
representing a true range in possible geological scenarios.
GRADE MODELLING
Mineralisation grades add another level of uncertainty of potential
mineralisation and have material impact on strategic scheduling
and blending studies. Statistics such as grade histograms and
variograms are difficult to infer from sparse data and are therefore
uncertain. Grade uncertainty ranges are addressed using a
bootstrapping technique for resampling histograms. Variogram
uncertainty was investigated, but was found to have only a minor
impact on the global grade tonnage distribution. Previous work on
the topic includes Berckmans and Armstrong (1999). They report
on a related but global approach for grade-tonnage estimation
accounting for grade uncertainty. The approach presented here
aims at local grade modelling, so grade estimation will be
reviewed in the next few paragraphs before presenting the
approach in the following section.
Fit-for-purpose grade estimation has to consider the fact that ore
is mined out as selective mining units (SMUs). The SMU
describes a defined volume that can be selected during mining and
is primarily linked to mining equipment. Typically, SMUs are
vastly larger than drill hole samples (eg Journel and Kyriakidis,
2004), but in relation to exploration drill hole spacing, they are
small. When data is sparse, linear in situ resource estimation likely
results in over-smoothed grade-tonnage curves. Such models
represent a too-low selectivity equivalent to excessively large
SMU size. A non-linear recoverable resource estimation approach
(or alternatively, a simulation approach) is required to enable
unbiased grade-tonnage curve estimation. The recoverable
resource models that are generated form the basis of a realistic
prediction of minable reserves that take dilution/ore loss,
economic factors, mining policies, etc into account.
Two approaches for recoverable resource estimation are
traditionally considered in the literature:
1. Derivation of the sample support distribution (pdf) for a
panel and application of a change of support correction, ie
conversion of sample support histogram to SMU support
histogram. Methods include uniform conditioning (UC) and
multiple-indicator kriging (MIK) combined with support
correction (eg Rivoirard, 1994; De-Vitry, Vann and
Arvidson, 2007).
46 Spectrum Series Volume 17
2. Conditional Simulation (CS) of geological and chemical
components of an orebody on a fine grid of sample support
values that are reblocked (averaged) to achieve SMU
support (Journel and Kyriakidis, 2004; Journel and
Huijbregts, 1977).
UC is based on the assumptions of the underlying Discrete
Gaussian Model (DGM). The DGM is a method in which the
original distribution is transformed to a standard normal
distribution using Hermite polynomials. This type of polynomial
has the remarkable property of being ‘orthogonal’ for Gaussian
values. This means that the transformation of the distribution for
a variance correction can be simply made by applying a
correction factor to the Hermite coefficients. Thus, there is no
assumption of normality or log-normality in the original data.
There is a reliance on the bi-normality of the transformed data
and expected SMU values. In UC, the local panel’s mean and
variance are estimated using kriging, while all other distribution
moments are based on the global data distribution and honour
Cartier’s relationship. Hence the proportional effect is poorly
captured by traditional UC. The method is fast and practical – a
multivariate implementation was recently provided (Deraisme,
Rivoirard and Carrasco, 2008). This approach enables a change
of support for a primary variable and a set of secondary
variables. The full pdf, F(X), of a primary variable, X, is
available for reporting at various cut-offs. Secondary variables,
Yi, can be reported as the expectation value of the conditional
pdf, E[F(Yi|X)].
Additional to the assumptions of the univariate DGM, the
multivariate DGM assumes multi-Gaussian behaviour of the
variables considered. MIK is a method with little assumptions
regarding the shape of the grade distributions. The estimation of
the point-support pdf is local, with variance and higher moments
derived by kriging from the search neighbourhood. The
application requires extensive variography even for the univariate
case. Various support correction methods are available to infer
local SMU support pdfs. Neither MIK nor UC incorporate global
or local uncertainty.
The common and commercially available CS techniques are
based on second order moments and also make assumptions
regarding the higher moments of the distributions. The
multi-Gaussian model is most widely used in the industry and is
fully defined by the Gaussian transformation of the sample
histogram and the related variogram model. An acceptable set of
CS realisations should reproduce those two statistics within
ergodic fluctuations (eg Goovaerts, 1997). Some multivariate CS
algorithms suffer from some problems, for example variance
reproduction (Babak and Deutsch, 2008). CS models capture
grade variability and local uncertainty, enabling their quantitative
evaluation within the CS model framework. These qualities cannot
be delivered by estimation techniques. In general, CS has
additional computation, storage and people requirements that raise
the question of cost/benefit balance. New efficient algorithms
retain SMU grades only and enable some CPU and storage savings
(eg Boucher and Dimitrakopoulos, 2009; Deraisme, Rivoirard and
Carrasco, 2008). However in this regional context model size
remains challenging even in terms of SMUs (eg for a 1 B tonnes
target – 5 M SMUs × 5 variables × 4 bytes × 50 realisations =
5 GB).
LOCAL CHANGE OF SUPPORT
The proposed approach local change of support, or LoCoS,
stands in between UC, MIK and CS and combines the strengths
of the three methods while avoiding some of their practical
limitations in this large scale context. Given a panel within a
domain, three steps are performed:
1. kriging – to derive a local weighted histogram,
2. spatial bootstrap of the local histogram, and
Advances in Orebody Modelling and Strategic Mine Planning I
 REGIONAL THREE-DIMENSIONAL MODELLING OF IRON ORE EXPLORATION TARGETS
3. change of support for each histogram realisation from the
second step.
Figure 4 illustrates the high level process flow for the LoCoS
grade modelling, showing decision points that occur when data
quantity is challenged, ie what to do if there is insufficient data
from which the distributions and spatial continuity can be
reasonably characterised. If data is too sparse, the local approach
suggests more information than is available and becomes
meaningless. Global assessment of grades and grade uncertainty is
more appropriate under such conditions. Given the regional
geological understanding and extensive history of iron ore
exploration and mining in the Pilbara, it is expected that analogous
deposits can be sourced to provide parameters (eg global pdf and
variogram models), even in the most data challenged situations.
Select panel
Sufficient No
local data?
Sufficient No
Yes global
Kriging for local pdf data?
Data analogue
Yes
Spatial bootstrap Build global pdf
Change-of-Support
Build parcel model
FIG 4 - Process flow diagram for the grade modelling approach.
Kriging
The kriging provides weights for the samples in the search
neighbourhood. Ordinary Kriging (OK) weights, accompanied
by primary and secondary assay variables, define an
experimental multivariate sample-support distribution for the
panel (Figure 5). Negative OK weights are set to zero and the
remaining weights restandardised to sum to 1. The search
criteria, such as minimum required number of samples, have to
balance the resolution of the final pdf and the potential for
OK for panel
ë1
ë2
ë3
FIG 5 - Ordinary kriging for panel gives weights λi for samples si in neighbourhood to derive local weighted histogram.
Advances in Orebody Modelling and Strategic Mine Planning I
negative kriging weights. Negative weights are tracked to enable
validation of the estimation performance. The locations of the
samples in the local neighbourhood define data sparsity and are
utilised to describe data redundancy in the spatial bootstrapping
process.
Spatial bootstrap
The randomisation is performed to account for the grade
histogram uncertainty given sparse data, ie the fact that sparse
data contains limited information regarding the true local pdf.
This step is performed using a spatial bootstrap method based on
the Gaussian transformation of the data pdf and unconditional
simulations of a Gaussian random field at the sample locations
(Deutsch, 2007). While classical bootstrapping assumes that data
is independent from one another, the spatial bootstrap method
accounts for the auto-correlation of spatial data. As two
auto-correlated samples contain information regarding each
other’s value, there is effectively less independent data when the
data is auto-correlated.
The spatial bootstrap method was adjusted to the requirements
of assessing local multivariate pdf uncertainty:
• The bootstrap is based on the primary variable and captures
data redundancy due to spatial correlation of this variable.
• The original method is univariate. For this approach, the
complete sample records are carried through to the resampled
data set to generate multivariate realisations. The complete
sample records maintain the experimental multivariate
relationships of the data.
• The realisations of the pdf are ranked based on data statistics
such as the mean grade of the primary variable – grade
models capturing bias scenarios can be generated from
equally ranking local pdfs.
The accumulation of uncertainty from multiple equally
ranking local pdfs was found to quickly exceed uncertainty
ranges based on the global data. To keep grade ranges reasonable
and aligned with the ranges specified in the case matrix the
number of realisations per panel can be kept relatively small.
Typically about ten realisations for each panel accumulate to an
uncertainty range comparable to 1000 global realisations (Figure 6).
Change of support
After the (optional) kriging and spatial bootstrap, a weighted
point-support multivariate pdf is available for the panel, with
multiple pdf realisations capturing local grade uncertainty. These
point-support panel pdf realisations require change of support
correction to derive the SMU-support grade tonnage distribution
pdf for panel
100
ë4
ë5
ë5 ë1
P
ë4
ë2
ë3
s3 s2 s1 s5 s4
Spectrum Series Volume 17 47
V OSTERHOLT, O HEROD and H ARVIDSON
FIG 6 - Cumulative histogram of Min, Mid and Max bootstrap
scenarios versus data for Al2O3 – based on 500 realisations
(global bootstrap) and ten realisations per panel (local bootstrap).
for the panel. The change of support is based on the multivariate
discrete Gaussian model (reviewed above – Deraisme, Rivoirard
and Carrasco, 2008). In the LoCoS algorithm, a Gaussian
anamorphosis is performed on each of the pdf realisations that
were obtained by the spatial bootstrap process, hence reducing
the strict assumption about the global distribution required in
traditional UC. This results in better representation of
proportional effect by LoCoS. Then the change of support is
performed with variance correction factors derived from the
LMC.
RESULTS
The process generates a set of 3D models reflecting the range of
grade-tonnage uncertainty for the model area. Figure 7 shows
three scenarios generated for an area of approximately 25 cubic
kilometres. The continuity of mineralisation (coloured on brown
topography) increases from the pessimistic to the optimistic
scenarios. The potty volumes around some drill holes are not
ideal, but the resulting global tonnage ranges are fit-for-purpose.
The kriging of panel grades can be evaluated using the same
statistical and graphical tools as any kriged estimate. Figure 8
shows the Eastings swath plots for Fe of blocked sample data
(red) and estimated blocks (blue). Overall grade fluctuations are
captured in the model with some spatial smoothing.
Figure 9 shows the resulting scatter plot of Fe versus Al2O3 for
SMU support (black) versus samples (pink). The red points
Pessimistic Best Guess
FIG 7 - Example of three mineralisation scenarios in plan view.
48 Spectrum Series Volume 17
display the model averages for each bin. The variance of both
variables is reduced in the SMU support with overall
reproduction of the conditional mean of Fe. Some grade parcels
are generated outside of the point scatter due to minor local
difficulties with the change of support parameters. Figure 10
summarises the grade-tonnage curves from nine model scenarios
showing three clusters in the tonnage and the grade plots
identifying the different geology scenarios. Grade uncertainty is
considerable, but mostly driven by geology and the associated
declustering effect. Grade uncertainty is a major contributor to
tonnage uncertainty at high Fe cut-offs.
CONCLUSIONS
The modelling process allows rapid generation of alternative 3D
models of exploration scenarios. The process is data-based and
transparently includes subjective expert judgements where
required. Implicit geological modelling tools enable the
integration of a range of geological data types and the quick
development of multiple lithological models. By basing the
estimation of potential mineralisation indicators on these models,
the resulting volumes are consistent with drill hole data, mapping
and the structural framework. The grade modelling approach
presented delivers multivariate models of recoverable resources
that reflect data statistics, including spatial trends. These models
are a realistic representation of deposit to regional scale grade
tonnage curves and avoid excessive smoothing, the typical
shortcoming of linear interpolators where data is sparse. Uni- and
bi-variate data statistics are well captured with the proposed
LoCoS method.
Due to the bi-variate nature of the change of support approach,
applying cut-offs is limited to a single variable. Using any
secondary cut-off variable would mean changing the selectivity for
that variable, as parcels with variable tonnages had to be selected,
with volumes generally exceeding the SMU size. In practice, this
demands an upfront decision regarding the main grade control
variable. An alternative approach could be formulated as a
variation of uniform conditioning (UC), with global uncertainty
being captured by utilising multiple Gaussian anamorphoses. This
approach would be easy to implement with UC with only a single
step additionally performed, the spatial bootstrapping of the global
data. However, the assumption of a known (estimated) local panel
grade would result in understating grade and tonnage uncertainty.
The 3D geology models add value by enabling visualisation of the
scenarios, thus developing further learning. They also have current
application in planning of exploration programs. The exploration
target model scenarios allow global grade tonnage reporting and
strategic planning. Future anticipated uses of either the current
geology models, or the geological modelling workflow, include
the wider mining related geosciences, such as geotechnical
engineering and hydrogeology.
Optimistic
Advances in Orebody Modelling and Strategic Mine Planning I
 REGIONAL THREE-DIMENSIONAL MODELLING OF IRON ORE EXPLORATION TARGETS
FIG 8 - Eastings swath plot validation for mean panel Fe grades.
FIG 9 - Bivariate validation of Fe versus Al2O3; pink – data, black –
SMU, red – average SMU.
FIG 10 - Grade tonnage curves for nine model scenarios.
Advances in Orebody Modelling and Strategic Mine Planning I
ACKNOWLEDGEMENTS
The authors wish to thank an anonymous peer reviewer for
constructive review of this paper, Sia Khosrowshahi (Golder
Associates) for his help in developing ideas for the grade
modelling approach and Richard Mattsson (BHP Billiton) for
practical guidance and motivation. Also thanks to BHP Billiton
Iron Ore for permission to publish this work.
REFERENCES
Babak, O and Deutsch, C V, 2009. Collocated cokriging based on
merged secondary attributes, Mathematical Geosciences, volume
41(8)921-926.
Berckmans, A and Armstrong, M, 1999. Bootstrapping: A fast way to
simulate QTz curves, Mathematical Geology, 31(4):471-485.
Boucher, A and Dimitrakopoulos, R, 2009. Block simulation of multiple
correlated variables. Mathematical Geosciences, 41(2), pp 215-237.
Calcagno, P, Chiles, J P, Courrioux, G and Guillen, A, 2008. Geological
modelling from field data and geological knowledge: Part I,
Modelling method coupling 3D potential-field interpolation and
geological rules, Physics of the Earth and Planetary Interiors,
171(1-4):147-157.
Caumon, G and Caers, J K, 2004. Assessment of global uncertainty for
early appraisal of hydrocarbon fields, Society of Petroleum
Engineers, paper 89943.
Deraisme, J, Rivoirard, J and Carrasco, P, 2008. Multivariate uniform
conditioning and block simulations with discrete gaussian model:
Application to Chuquicamata deposit, in Proceedings Eighth
International Geostatistics Congress (eds: J M Ortiz and X Emery),
pp 69-78.
Deutsch, C V, 2007. A statistical resampling program for correlated data:
Spatial_boostrap, paper 401, in Annual Report Centre for
Computational Geostatistics, Report 6.
De-Vitry, C, Vann, J and Arvidson, H, 2007. A guide to selecting the
optimal method of resource estimation for multivariate iron ore
deposits, in Proceedings Iron Ore 2007, pp 67-78 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
Applied Geostatistics Series, 483 p (Oxford University Press:
New York).
Journel, A G and Huijbregts, C J, 1977. Mining Geostatistics, 600 p
(Blackburn Press: Caldwell).
Journel, A G and Kyriakidis, P C, 2004. Evaluation of Mineral Reserves:
A Simulation Approach, 216 p (Oxford University Press: New York).
Levy, B and Mallet, J L, 1999. Discrete smooth interpolation: Constrained
discrete fairing for arbitrary meshes [online], GoCAD whitepaper.
Available from: <http://alice.loria.fr/publications/papers/1999/smooth
ing/smoothing.pdf>
Mallet, J L, 1992. Discrete smooth interpolation, Computer Aided Design
Journal, 24(4):263-270.
Nowak, M and Verly G, 2007. A practical process for simulation, with
emphasis on gaussian simulation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 69-78 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Rivoirard, J, 1994. Introduction to Disjunctive Kriging and Non-Linear
Geostatistics, 180 p (Clarendon Press: Oxford).
Spectrum Series Volume 17 49

Conditional Simulation of Grades Controlled by Geological
Indicators
C Xu1 and P A Dowd2
ABSTRACT
This paper describes a method of conditionally simulating an orebody in
which geological indicators have a controlling influence on the
occurrence and spatial distribution of the grades of primary economic
variable. The simulation comprises two components: 1. grades, and 2.
geological controls.
In the example provided in the paper, the authors use plurigaussian
simulation to simulate the occurrence of quartz veins and sequential
Gaussian simulation to simulate the associated gold grades. The
simulated orebody can then be sampled at various scales to yield
estimates of gold content and associated geological indicators. Estimates
derived from the data at each sampling scale can be compared against the
complete simulated orebody to assess the effects of estimated indicators
on recoverability. The cost of additional sampling/drilling can be assessed
against the increased profit generated by more accurate prediction of the
geological indicators.
INTRODUCTION
Geological indicators are an important element in the selective
open pit mining of many types of orebodies. A good example is
provided by quartz-vein controlled gold orebodies in which
selective mining units are defined by outlines around visible
occurrences of quartz veins and/or natural fractures on exposed
benches. Selection can be implemented in this way only when a
mining bench is exposed and even then only exposed veins and
fractures can be used in selection. It is much more difficult to
include the effects of such selection processes in recoverable
reserve estimates. The ability to predict accurately the
occurrence of geological indicators and their effects on ore
boundaries largely determines the amount of ore loss and ore
dilution; minimising each of these is a major determinant of
profitability.
Problems associated with identifying geological indicators and
their relationship to ore zones are exacerbated in low-grade gold
orebodies. This is especially so when the only samples available
are from percussion drilling, from which it is not possible to
discern the detailed nature of the geological indicator.
THE OREBODY
A full description of the geology and mining operation can be
found in Dowd (1995). Gold mineralisation occurs within a
network of steeply dipping quartz veins in the contact between
older granodiorite and limestone/acid volcanic rocks. The gold
has an erratic distribution and is mainly concentrated in and
around high-grade quartz veins. Granitic zones between the
high-grade veins contain fine-grained gold associated with
pyrite.
The geological relationship between quartz and gold is a
fundamental element in the selective mining of ore and the final
delineation of ore zones is achieved largely by mapping quartz
veins on exposed benches. This relationship is, thus, an
1. Senior Lecturer, School of Civil, Environmental and Mining
Engineering, Engineering North Building, The University of
Adelaide, SA 5005. Email: chaoshui.xu@adelaide.edu.au
2. FAusIMM, Executive Dean, Faculty of Engineering, Computer and
Mathematical Sciences, The University of Adelaide, SA 5005.
Email: peter.dowd@adelaide.edu.au
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
important controlling factor in ore reserve estimation and
selective mining.
Exploration and development drilling was originally on a
30 m × 30 m grid with some areas drilled at 30 m × 15 m and
peripheral areas drilled at 60 m × 60 m, one axis of the grid
having a parallel orientation to the major strike direction.
Geostatistical studies indicated an optimal drilling grid of
25 m × 12 m, but a cost-imposed compromise increased the
minimum dimensions to 30 m × 15 m. The holes were drilled
from the surface at a dip of approximately 40° to the north or
south to intersect the predominantly vertical quartz veins at as
great an angle as possible, as illustrated schematically in Figure 1.
Initial drilling was by direct circulation and thereafter by reverse
circulation. Hole diameters varied from 90 mm to 150 mm and
were eventually standardised at 135 mm.
Relationship between quartz and gold
The correlation coefficient between the quartz and gold contents
of samples is not significant even though there is strong
geological and mineralogical evidence for the association of gold
with quartz veins. This lack of statistical correlation may be
caused by the confinement of the association to the contact zones
between quartz veins and granodiorite. If this is the case, the
important factor would be the number of such contacts or veins
within a given sample rather than the total amount of quartz in a
sample, as illustrated in Figure 2. This information is not
available from percussion holes. Another explanation may be
that, although gold is associated with quartz, the mineralisation
may have independent controls within the quartz-vein system.
Note, however, that there is no difference in the spatial variability
of gold directly associated with quartz and that of gold not
directly associated with quartz as shown by the two downhole
variograms in Figure 3. These variograms can be modelled as
spherical models with a nugget variance and two structures – the
first with a range of 4 m and the second with a range of 20 m.
The complete three-dimensional model is given in Table 3.
Figure 4 presents log-probability plots of the cumulative
histograms of gold assays of samples within a representative
zone of the orebody. The Figure shows the cumulative histogram
of the gold assays of all samples with corresponding quartz
assays of zero wt per cent together with the cumulative
histogram of gold assays of samples with corresponding quartz
assays greater than zero wt per cent. In spite of the geological
evidence, the only significant difference between the two
log-probability plots is the mean grade of the two sets of values.
The difference between the mean of the gold values associated
with quartz and of those not associated with quartz is largely due
to a few high-grade samples. A summary of statistics for the
representative zone is given in Table 1.
The histograms in Figure 4 and the statistics in Table 1
illustrate a typical problem in the evaluation of gold orebodies –
the presence of a small number of high-grade outliers. The
maximum grade in the data set that was used to calculate the
statistics in Table 1 was 4360 g/t. If the high-grade value of
4360 g/t is excluded from the calculation, the mean of the
quartz-associated gold assays is 3.45 g/t and the variance is
824.4 (g/t)2. The mean and variance can be further reduced by
excluding two or three other high-grade outliers.
Spectrum Series Volume 17 51
C XU and P A DOWD
FIG 1 - Schematic of typical cross-section.
quartz
(a) (b)
quartz
FIG 2 - Possible relationships between gold content and numbers
of quartz-granodiorite contacts. Examples of drill hole
intersections: (A) with low quartz content and high number of
quartz-granodiorite contacts and (B) with high quartz content and
low number of quartz-granodiorite contacts.
FIG 3 - Downhole variograms of the logarithms of gold grades of
samples in which no quartz is recorded and for samples in which
quartz is recorded.
It is possible to fit three lines of slightly different slopes at
values of 1.0, 0.01 and 0.001 g/t on the cumulative histograms in
Figure 4. The 1 g/t threshold appears to be related to gold-
bearing quartz-vein occurrences, although the relationship cannot
be quantified.
A further study was conducted to determine whether samples
with a significant gold assay and a zero quartz assay occur close
to the occurrence of quartz. A summary of the study is given in
Table 2, which was constructed by selecting all samples with a
quartz assay of zero wt per cent and then searching for the
nearest sample in the same drill hole with a non-zero quartz
52 Spectrum Series Volume 17
assay. As an example, consider the entry for gold assays greater
than or equal to 0.5 g/t. There are 644 samples with zero wt per
cent quartz and a gold assay greater than or equal to 0.5 g/t.
Of these, 39 per cent are within an average of 1 m of a non-zero
quartz assay, 58 per cent are within 2 m of a non-zero quartz
assay and 68 per cent are within 3 m of a non-zero quartz assay.
These distances are measured along the drill hole, so that if
quartz is encountered within 1 m of a sample, it occurs in the
contiguous sample on either or both sides. Distances of 1 m, 2 m
and 3 m along a drill hole are equivalent to horizontal distances
of 0.7 m, 1.4 m and 2.1 m, respectively.
The following observations can be made on the basis of the
information summarised in Table 2:
• a nearly constant 40 per cent of all significant gold assays
that are not associated with quartz occur within 1 m of a
quartz occurrence, ie 60 per cent are further than 1 m (0.7 m
horizontally) from a quartz vein;
• a nearly constant 60 per cent of all significant gold assays
that are not associated with quartz occur within 2 m of a
quartz occurrence, ie 40 per cent are further than 2 m (1.4 m
horizontally) from a quartz vein; and
• a nearly constant 70 per cent of all significant gold assays that
are not associated with quartz occur within 3 m of a quartz
occurrence, ie 30 per cent are further than 3 m (2.1 m
horizontally) from a quartz vein.
The mean gold grade of samples that have a zero quartz assay
and a significant gold assay and which are within 1 m of a quartz
vein (column 5 in Table 2) is consistently and significantly less
than the mean gold grade of all samples with zero quartz assay
(column 3).
The three observations appear to be evidence that significant
gold mineralisation is not directly associated with quartz veining
and/or that there is significant sample contamination, which is
often a problem in percussion drilling, especially with direct
circulation. If contamination is a factor, the nearest occurrence of
quartz for any sample should, more often than not, be further up
the hole. As there is no evidence of this, the only possible
conclusion is that, although the coarsest-grained particles of gold
may be associated with quartz contacts, there is a significant
proportion of finer-grained gold that is not directly associated
with quartz veins.
Extending the calculations shows that approximately 80 per
cent of all samples with zero wt per cent quartz and significant
gold grades are within 5 m of a quartz occurrence. A downhole
distance of 5 m corresponds to a horizontal distance of
approximately 3.5 m. When all samples are considered (ie
Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL SIMULATION OF GRADES CONTROLLED BY GEOLOGICAL INDICATORS

100

10

Gold grade, g/t, of 1 m drill hole samples

Quartz > 0% Quartz = 0%

1.0

0.1

0.01

0.001
0.01 0.05 0. 1 0.2 0.5 1 2 5 10 20 30 40 50 60 70 80 90 95 98 99 99.8 99.9 99.99

Cumulative frequency

FIG 4 - Log-probability plots of gold grades for samples containing no quartz and samples containing quartz.

TABLE 1 shown in Figure 5 in which a spherical model with a range of

Statistics of gold assays with and without associated quartz
(subvolume).
Statistic Quartz = 0 wt%
(n = 3849)
Mean, g/t
Variance, (g/t)²
Median, g/t
25th percentile, g/t
75th percentile, g/t
90th percentile, g/t
including samples that contain quartz) 72 per cent of all gold
assays greater than or equal to 0.5 g/t are either directly associated
with quartz or are within 1 m of a quartz occurrence; 80 per cent
are either directly associated with quartz or are within 2 m of a
quartz occurrence; and 85 per cent are either directly associated
with quartz or are within 3 m of a quartz occurrence. These
percentages increase as the cut-off grade increases because a
larger proportion of high-grade gold is associated with quartz.
These observations have important implications for selective
mining, in particular for optimum mining or selection widths.
The relationship of gold and quartz within distances of up to
3 m is reflected in the downhole variogram for quartz grades
3 m has been fitted. This range is very similar to the range of the
first structure observed on the downhole variograms for gold
(Figure 3 and Table 3) and may reflect the controlling influence
Quartz >0 wt% of quartz veins on gold occurrence.
(n = 1856) The challenge in the simulation of the gold grades is to
0.95 5.79 reproduce the observed associations with quartz in the absence of
any significant statistical correlation and any spatial covariation
55.3 11 050.1
in the form of cross-variograms. Independent univariate
0.06 0.30 simulation of gold and quartz grades will reproduce the spatial
0.02 0.08 correlation of each variable but not the physical association
described above. Standard forms of co-simulation of gold and
0.25 1.30
quartz do not help either because of the negligible spatial
1.00 5.00 cross-correlation between the two variables. We propose a
modified version of sequential Gaussian simulation which uses
an adaptive local mean during the simulation. The adaptive local
mean in this case is controlled by previously simulated quartz
grades.
As a preliminary test for the proposed method, we use only the
eastern part of the orebody. In this subset, there is a total
of 16 626 samples of which 31 per cent (5077) record quartz
occurrence. The association between gold and quartz can be
demonstrated by plotting the average gold grade of samples
against the distance of the samples to the closest quartz sample,
as shown in Figure 6a. Note that the average grade for zero
distance is the average gold grade for those samples that contain
quartz. Average grades for other distances are the average gold
grades for samples in which quartz is not present and they are

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 53
C XU and P A DOWD

TABLE 2
Proportion of samples with 0 wt per cent quartz within specified distance of quartz veins.

Gold, > g/t No of samples Mean of One metre Mean of Two metre Mean of Three metre Mean of
samples, g/t distance per samples, g/t distance per samples, g/t distance per samples, g/t
cent of total cent of total cent of total
0 3849 0.95 31 0.98 49 1.1 62 1.14
0.2 1079 3.24 39 2.7 58 3.24 69 3.47
0.5 644 5.23 39 4.27 58 5.16 68 5.61
1 379 8.39 37 7.1 55 8.6 68 9.06
1.5 299 10.31 38 8.57 56 10.36 70 10.82
2 244 12.25 40 9.69 57 12.09 72 12.58
2.5 207 14.03 40 10.98 56 14.12 72 14.45
3 169 16.58 41 12.62 59 16.09 73 16.82
3.5 142 19.11 40 14.58 58 18.75 73 19.46
4 134 20.02 41 14.97 58 19.52 74 20.25
4.5 121 21.72 41 16.05 60 20.8 74 21.85
5 114 22.75 42 16.53 60 21.74 74 22.86
5.5 102 24.82 44 17.28 63 22.77 74 24.94
6 96 26.01 44 18.1 63 23.61 73 26.59
7 82 29.35 44 20.04 63 26.58 72 30.35
10 55 39.71 38 28.34 58 37.98 67 43.49

The average wt per cent quartz corresponding to Figure 6a is

TABLE 3 given in Figure 6b. The figure further illustrates the lack of
Spherical model variogram used in the sequential Gaussian correlation between the gold and quartz grades.
simulation.
In this application, a univariate gold simulation will reproduce
Nugget Structure 1 Structure 2 the spatial correlation, but not the association relationship, as
shown in Figure 7.
C 0.41 0.35 0.24
a (range) - strike 30 30
COMBINED PLURIGAUSSIAN AND SEQUENTIAL
- down-dip 25 25 GAUSSIAN SIMULATION
- cross-dip 4 20
As quartz vein location is a controlling factor in gold occurrence it
is sensible to include the quartz vein information when simulating
the gold grade. The approach proposed in this paper involves two
distinct steps. The quartz vein is simulated first using the
plurigaussian method, which provides the basic structure for the
next step in the simulation. In the second step the gold grades are
simulated using the sequential Gaussian method but using adaptive
local means rather than the global mean. The local mean for each
Gaussian simulated value is obtained as a function of the shortest
distance from the simulation point to a previously simulated quartz
vein in a manner similar to that illustrated in Figure 6a. On this
basis the sequential Gaussian simulation reproduces the spatial
correlation of gold grades and the adaptive local mean obtained
from the simulated quartz model reproduces the gold-quartz
association relationship shown in Figure 6a.
Plurigaussian simulation (Le Loc’h and Galli, 1996) is a
versatile method for simulating geological structures. It truncates
multiple correlated Gaussian fields to produce indicator
variables, which represent the geological units to be simulated.
For details of the method used here and the open-source codes
for the implementation, readers are referred to Dowd (2003) and
Xu et al (2006).
FIG 5 - Downhole experimental variogram and model for As the actual quartz grade is not an important indicator for
quartz grades. gold concentration, only presence/absence of quartz is of interest
in simulating the quartz vein model for the final simulation. The
indicator for the presence/absence of quartz can be treated as a
calculated on the basis of the distances of the samples to the category variable and the plurigaussian method can then be used
closest occurrence of quartz. The relationship shown in Figure 6a to simulate it. Plurigaussian simulation in this case is a truncated
applies to the subset used in this study and could be replaced by univariate Gaussian simulation, as only one Gaussian variable is
a relationship relevant to an application. required to simulate one indicator variable.

54 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL SIMULATION OF GRADES CONTROLLED BY GEOLOGICAL INDICATORS

Average Au (g/t) Average Qz (%)

20
1.6
16
1.2
12
0.8
8

0.4 4

0 0
0 10 20 30 0 10 20 30
(A) Distance to Qz vein (m) (B) Distance of Au sample to Qz vein (m)
FIG 6 - Average gold and quartz grades versus the gold sample distance to nearest quartz occurrence.

Average Au (g/t) a standard Gaussian variable corresponding to a cumulative

frequency of 0.309). Two variables are created to define the
1.6 spatial correlation of quartz veins for the plurigaussian
Traditional sgsim simulation. F1 represents the presence of quartz (ie F1 = 1 if
1.2 quartz >0 per cent and 0 otherwise) and F2 represents the
absence of quartz (ie F2 = 1 if quartz = 0 per cent, 0 otherwise).
0.8 The variogram of F1 and the cross-variogram F1-F2 are shown in
Figure 8. These variograms are used by PGSim1 (Xu et al, 2006)
0.4 to derive the spatial correlation model for the Gaussian field. The
optimal model found is an exponential model with a = 17 m.
0 This Gaussian model will produce the theoretical spatial
0 20 40 60 correlations for F1 and F2 shown as the solid lines in Figure 9.
Distance to the closest Qz sample (m)
Distance to the closest Qz sample (m)
Sequential Gaussian model
FIG 7 - Result from traditional SGSIM (Deutsch and
Journel, 1992). The sequential Gaussian method is used for the simulation of
gold grades. The variogram model for the gold grades is reported
Monte Carlo sampling is used to derive the adaptive local mean in Dowd (1995). The parameters of the model fitted to the
corresponding variogram of the normal score transform of the
to be used in the simulation of gold grades. The distribution used
gold grades are given in Table 3. This variogram model for the
for the Monte Carlo simulation is derived directly from the data strike-plunge direction is shown in Figure 10 together with the
set. For example, Figure 8 shows the histogram, derived from the experimental variogram calculated from the data and the
dataset, of the gold grades of samples that are 2 m away from the variogram calculated from a simulation.
nearest sample containing quartz. This histogram is used to
generate the adaptive local mean for the gold grade simulation at
those simulation locations that are 2 m away from the nearest SIMULATION RESULTS
simulated quartz occurrence. Figure 11 shows a simulated quartz vein model generated by the
plurigaussian simulation. This model is used, as described above,
No. Samples( 1000) to derive the adaptive local mean for the sequential Gaussian
4 simulation of the gold grades. A cross-section from one such
simulation is shown in Figure 12.
3 The spatial correlation models are reproduced satisfactorily by
the combined plurigaussian and sequential Gaussian simulation,
2 as is evident from Figures 9 and 10. Most encouragingly, by
using the approach described here, the reproduction of the
association between quartz and gold has been improved, as
1
shown in Figure 13. The improvement over models generated by
sequential Gaussian simulation without the control imposed by
0 the prior quartz simulation (Figure 7) is significant.
0.01 0.51 1.01 1.50 2.00 2.50
Histogram of A× grade for distance = 2 m CONCLUSIONS
FIG 8 - Gold grade histogram derived from data. This paper describes a simple two-step simulation method to
incorporate the controlling effects of quartz veins when there is
no significant direct correlation between gold and quartz. The
SPATIAL CORRELATION MODELS geological structure in this approach is simulated by the
plurigaussian method and the primary variable (gold) is
Plurigaussian model simulated by the sequential Gaussian method with an adaptive
local mean derived from the simulated geological structure. The
The proportion of samples in the dataset that contain quartz (ie improvement in terms of the reproduction of the association is
samples with quartz grade > zero per cent) is 0.309. This number significant. If diamond drill core data are available and logged
is used in the plurigaussian algorithm to determine the truncation it would be possible to simulate the detailed shapes of the veins
threshold t = -0.498 for the Gaussian variable (ie t is the value of rather than the simplistic simulation of quartz occurrence.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 55
C XU and P A DOWD

γF1 γF1-F2 Distance (m)

0.25 0
0 100 200 300
0.2 -0.05

0.15 -0.1

0.1 -0.15

0.05 -0.2
Distance (m)
0 -0.25
0 100 200 300
Data Model Simulation

FIG 9 - Spatial correlation models for F1 and F2.

γ - normal scores
1.2
Data
1
Model
0.8
Simulation
0.6
0.4

0.2
Distance (m)
0
0 20 40 60 80 100

FIG 10 - Variogram of normal scores in the strike-plunge direction.

Quartz vein

North-south cross section @ easting=1808

3D quartz model within the defined orebody

FIG 11 - A simulated quartz model generated by plurigaussian simulation.

North-south cross-section @ easting = 1808

0.0 >0.5
FIG 12 - A simulation of gold grades generated by sequential Gaussian simulation.

56 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL SIMULATION OF GRADES CONTROLLED BY GEOLOGICAL INDICATORS

Average Au (g/t) The work described in this paper is only preliminary and is
presented primarily as a test of the proposed method. The
1.6 encouraging results reported warrant further detailed
investigation for the entire orebody. More detailed modelling of
1.2 the plurigaussian model is also necessary.
One simulation
0.8 REFERENCES
0.4 Deutsch, C and Journel, A G, 1992. Geostatistical Software Library and
User’s Guide, 340 p (Oxford University Press).
0 Dowd, P A, 1995. Björkdal gold-mining project, northern Sweden,
Transactions of the Institutions of Mining and Metallurgy Mining
0 10 20 30
Industry, Mining Technology, 104:A149-A163.
Distance to the closest simulated Qz vein Dowd, P A, Pardo-Igúzquiza, E and Xu, C, 2003. PLURIGAU: A
computer program for simulating spatial facies using the truncated
plurigaussian method, Computers and Geosciences, 29(2):123-141.
FIG 13 - Results from the two-step simulation.
Le Loc’h, G and Galli, A, 1996. Truncated plurigaussian method:
Theoretical and practical points of view in Proceedings Fifth
One significant application of the simulations generated by International Geostatistics Congress (eds: E Y Baafi and N A
these methods is the optimal design of drilling grids to detect Schoffield), 1:211-222 (Kluwer Academic Publishers: Dordrecht).
geological indicators and to incorporate the effects of these Xu, C, Dowd, P A, Mardia, K V and Fowell, R J, 2006. A flexible true
indicators in recoverable reserve estimates. plurigaussian code for spatial facies simulations, Computers and
Geosciences, 32(10):1629-1645.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 57

Mining Geostatistics with the Open Source SGeMS Software and
the GsTL Library
A Boucher1
ABSTRACT
Stochastic representations of mineral deposits are complex and off-the
-shelf algorithms are not always readily suitable for their modelling. The
free and open source software SGeMS (Stanford Geostatistical Modelling
Software) and the library GsTL (Geostatistical Template Library) with its
collection of advanced probabilistic algorithms, provides convenient
resources for deposit and uncertainty modelling, research and algorithm
development. SGeMS also encompasses the open source Python scripting
language that allows further processing of geostatistical simulations such as
grade control. Finally, the latest multiple-point simulation algorithms are
also available in the software package.
This paper explores the capabilities of the libraries for mining
geostatistics. Two examples are used to illustrate the potential of the
SGeMS software in developing practical tools. First, a C++ plug-in that
seamlessly integrates the Group Sequential Gaussian Simulation (g-sgsim)
into the software SGeMS is presented. Then a Python script is written to
select optimal variogram parameters through a cross-validation procedure.
INTRODUCTION
Numerical 3D representations of mineral deposits with geo-
statistical techniques are widely used both in academia and in
industry. The use of geostatistical modelling techniques in mining
applications is only possible through efficient and adaptable
software. Existing software specially designed for the mining
industry has made the spread of geostatistical techniques in
modelling and engineering possible. However, mineral deposits
and other natural phenomena can be so complex that algorithms
offered in off-the-shelf software may not be suitable without
significant modifications. There are two solutions:
1. either a software package offers an exhaustive series of
algorithms, or
2. it allows the user to have access to the code and add/modify
algorithms.
For instance, the availability of the source code is a significant
contributing factor for the popularity of the GSLIB suite of
algorithms (Deutsch and Journel, 1998).
This paper shows the capabilities of SGeMS (Remy, Boucher
and Wu, 2009), an open source geostatistics software, for mining
applications. SGeMS, fully written in C++, includes geostatistical
algorithms developed at the Stanford Center for Reservoir
Forecasting (SCRF) at Stanford University. As of October 2008, it
has been downloaded more that 12 000 times (not necessarily
unique users). It was written by Remy (Remy et al, 2002; Remy,
2004a, 2004b), and is maintained by a group of developers at
Stanford University and elsewhere. SGeMS was written with two
goals in mind. The first goal, geared toward the end-user, was to
provide a user-friendly software which offers a wise selection of
geo- statistics tools – ranging from commonly used variogram
based algorithm (eg kriging, cokriging, indicator kriging,
sequential Gaussian simulation and cosimulation, indicator
sequential simulation and direct sequential simulation), to the
more recent training image-based (multiple-point statistics)
simulation (Strebelle, 2002; Wu, Boucher and Zhang, 2008;
Journel, 2007). The user friendliness of SGeMS comes in part
1. Environmental Earth System Science, Stanford University, Yang and
Yamazaki Environment and Energy Building 4215, 473 Via Ortega,
Room 353, Stanford University, Stanford CA 94305, USA.
Email: aboucher@stanford.edu
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
from its non-obtrusive graphical user interface and from the
possibility to directly visualise data sets and results in a full 3D
interactive environment. The second objective was to design
software that would cater to the needs of power-users. In SGeMS,
most of the operations performed using the graphical interface can
also be executed through scripts. The integrated support for the
Python scripting language enables the creation of macros for
simple tasks all the way to complete nested applications with an
independent graphical interface. New features can be conveniently
added into SGeMS through a system of plug-ins, ie pieces of
software which can not be run by themselves, but complement a
main software. In SGeMS, plug-ins can be used to add new
geostatistics tools or utilities such as defining new file filters to
import/export data into appropriate file format.
THE CASE FOR OPEN SOURCE SOFTWARE
SGeMS is free and open source. Open source software gives the
users access to the source code, which can be modified with or
without some limited restrictions, depending on the license
requirement. For example, SGeMS is distributed under a GNU
General Public License (GPL) and must always be distributed with
its source code. Any changes or additions to the software are
automatically under the GPL license and the source code must be
provided whenever it is distributed. See http://www.gnu.org/copy
left/gpl.html for more details on the GPL license. Open source
software gives flexibility and freedom to the users. The source
code can be modified to address complex problems that may not
have been easily handled from the traditional interface of
geostatistical algorithms. Additionally, open source software can
grow through contributions from the user’s community. Access
to the source code also allows the users to track and correct bugs
without waiting for an official release or patches from a vendor.
More importantly, the transparency permits the user to validate
the algorithms by inspecting the actual codes instead of using
them as black boxes.
ADDING A GEOSTATISTICAL ALGORITHM TO
SGeMS
New algorithms can be seamlessly integrated in SGeMS through a
plug-in mechanism. Once loaded, all user-generated algorithms
share the same interface with SGeMS as the existing ones, and
have access to the same facilities that SGeMS provides, such as
visualisation and data analysis. To facilitate the writing of
algorithms, the user has access to all the libraries of SGeMS (such
as the search neighbourhood and sequential simulation) and of the
Geostatistical Template Library (GsTL), which contains geo-
statistical functions and utilities such as probability distribution
functions, builders of kriging systems and variograms/covariances
functions. Thanks to the generic implementation of these libraries,
the user can focus on the specificity of the new algorithm without
having to rewrite basic functions that are common to most geo-
statistical algorithms. A SGeMS plug-in is composed of two
elements:
1. a shared library (‘.so’ file on Linux, ‘.dll’ file on Windows),
and
2. a description of the graphical interface used to prompt for
parameters (‘.ui’ file).
Spectrum Series Volume 17 59
A BOUCHER
The shared library is obtained by compiling the source code,
and the user interface is created with the Qt Designer software
(distributed with SGeMS). Writing a SGeMS plug-in requires
moderate knowledge of the C++ language – a college level
introductory course is sufficient. SGeMS uses object-oriented
programming and template programming and modifying an
existing algorithm is the easiest and most efficient strategy to
create new algorithms.
The first step in writing a plug-in is to define a new class
derived from the predefined class Geostat_algo and to redefine
the three virtual functions:
1. initialise,
2. execute, and
3. name.
The purpose of these functions follows:
• Initialize() – Parse the input from the user and initialise the
parameters used by the execute function. These parameters
are usually stored as member variables of the algorithm class.
• Execute() – Runs the algorithm using the parameters
recorded with the initialise() function.
• Name() – Return the name of the algorithm.
• Graphical user interface (‘.ui’ file) – Create a user interface
to record the user parameters – it is automatically loaded
with the plug-in.
Any algorithm derived from Geostat_algo can be seamlessly
integrated into the SGeMS platform. A simple example would be
a wrapper function that takes the user inputs, reformats it into a
parameter file and calls an external program, for instance an
existing FORTRAN code. Once compiled, the plug-in (.dll or
.so) should to be copied in the plugin/Geostat/ folder where
SGeMS is installed. When SGeMS is launched, all shared
libraries in that folder are loaded and they appear in the list of
available algorithms.
PLUG-IN EXAMPLE – THE GENERALISED
SEQUENTIAL GAUSSIAN SIMULATION
The generalised sequential Gaussian Simulation (g-sgsim)
(Dimitrakopoulos and Luo, 2004) is a generalisation of the
sequential Gaussian simulation (sgsim) where instead of
simulating one point at a time with a simple kriging system, a
group of points are simultaneously simulated with a LU
decomposition algorithm (Davis, 1987). The g-sgsim algorithm
is faster than the traditional sgsim algorithm, an important aspect
when simulated realisations of deposits are large, for example,
tens of millions of nodes (Benndorf and Dimitrakopoulos, 2007).
The multivariate probability distribution function can be
written in term of Np -points group decomposition:
Ν1 N1 + N p = N2
f ( u1 ,..., uN ; z1 ,..., zN Ω 0 ) = Π f ( ui ; zi Ω i − 1 ).
i=1
Nv− 1 + N p = Nv
Π f ( ui ; zi Ωi−1 )
i = Nv− 1 + 1
where:
Ω0 is the set containing the observed (sampled) data
Ω i − 1 is the set of the conditioning data when simulating the ith
point
The domain D to be simulated is partitioned into Nv groups of
Np points. For example, each group may identify a mining block.
In the case where Np = 1, g-sgsim is the same algorithm as sgsim.
Let Z(u) be a Gaussian RF with mean mz and covariance model
60
Π f ( ui ; zi Ω i − 1 ).
i = N1 + 1
Spectrum Series Volume 17
CZ(h). The covariance matrices CII, CpI and Cpp, respectively the
covariance between the conditioning nodes, between the group to
be simulated and the conditioning nodes and between the nodes
of the group. These covariance matrices must be computed to
build the LU system of equation with zI as the data vector
contained in Ω i − 1 and zp the vector of the group of nodes to be
simulated:
⎡z I ⎤ ⎡L II 0 ⎤⎡wI ⎤
z = Lw = ⎢ ⎥ = ⎢ ⎥⎢ ⎥
⎣z p ⎦ ⎣L pl L pp ⎦⎣w p ⎦
where:
⎡C II C Tpl ⎤ ⎡L II 0 ⎤⎡ LTII LTpl ⎤
⎢ ⎥=⎢ ⎥⎢ ⎥
⎢⎣C p I C pp ⎥⎦ ⎣L pl L pp ⎦⎢⎣0 LTpp ⎥⎦
and:
wp is a vector of independent realisation from a standard
Gaussian distribution
The simulated group of nodes zp is then:
z p = C plC −1
II z 1 + L pp w p
A normal score transform may be needed before and after
simulation to transform the data to and from a Gaussian
distribution.
The g-sgsim algorithm can be summarised as follows:
• perform a normal score transform (if necessary), and
• initialise a random path visiting each group of pixels once.
For each group of pixels along the random path:
• get the neighbouring hard data and previously simulated
nodes,
• build the LU system with the neighbourhood data,
• simulate the group by multiplying the L matrix with a vector
of independent standardised Gaussian realisations,
• set the simulated points as data, and
• back-transform the simulated data from a Gaussian
distribution to its original distribution (if necessary).
From an algorithmic perspective the differences between
sgsim and g-sgsim are:
• the simulation path visits discontinuous groups of nodes
instead of individual nodes,
• the drawing is done with the LU decomposition instead of
simple kriging, and
• additional user input required about the size and geometry of
the group.
WRITING THE G-SGSIM PLUGIN
This section shows only the relevant parts of the C++ code that is
modified from the sgsim algorithm. In addition to the initialise
and execute functions, a new sequential simulation function that
is specifically written to handle group simulation is presented.
The lines starting with the ‘//’ characters indicate comments and
are not part of the actual code. The ‘…’ characters indicate code
that is not presented here for brevity.
Function initialize
The function initialize() is a slightly modified version of the one
in the sgsim algorithm. In the Gsgsim version, a group
simulation path and a group_cursor are initialised instead of a
regular simulation path. The group_cursor extracts the pixels
belonging to any given group along the simulation path.
Advances in Orebody Modelling and Strategic Mine Planning I
 MINING GEOSTATISTICS WITH THE OPEN SOURCE SGEMS SOFTWARE AND THE GSTL LIBRARY
Gsgsim::initialize (const Parameters_handler* parameters,
Error_messages_handler* errors ) {
...
// Instantiate the simulation path with the user-specified group
// (nx,ny,nz)
path_ = new Gsgs_sim_path(simul_grid_,GsTLTriplet[nx,ny,nz]);
// Instantiate an object that retrieves all the points within a
//group of size (nx,ny,nz)
group_cursor_ = new Group_sequential_cursor(simul_grid_,
GsTLTriplet[nx,ny,nz]);
// Instantiate and parameterise a neighbourhood class with user-
// specified ranges and angles.
neighborhood_ = SmartPtr<Neighborhood>(
simul_grid_->neighborhood( ranges, angles, &
covar_ ) );
...
} path_->begin(), path_->end(),
Function execute
The execute function generates nreal simulations specified by the
user and parameterised by the initialise function. The function
first performs the normal score transformation of the data and
initialises the LU simulator with the correct covariance. Then for
each simulation, it initialises a new random path and executes the
sequential simulation on the groups. The realisations are then
back-transformed from a Gaussian distribution to their original
distribution (if requested by the user).
Gsgsim::execute(GsTL_projects*) {
...
// Perform normal score transform (if necessary) using the non-
//parametric distribution object target_cdf_ (defined in GsTL) as
//distribution for the data.
if(use_target_hist_) {
FIG 1 - SGeMS interface with the algorithm panel (left) and the visualisation panel (right).
Advances in Orebody Modelling and Strategic Mine Planning I
harddata_property_ =
geostat_utils::gaussian_transform_property(
harddata_property_, target_cdf_, harddata_grid_ );
}
// Instantiate the object lu_sim to perform LU simulation with
// covariance covar_, the random number generator gen on the
// grid simul_grid_. The code to generate LU simulation is already
// available from the the GsTL library.
LU_Simulator lu_sim(covar_,gen,simul_grid_);
// Loop over all realisations
for( int nreal = 0; nreal < nb_of_realisations_ ; nreal ++ ) {
// Get a new group random path
path_->init_random_path();
// Perform the simulation from the first group (path_->begin()) to
// the last group (path_->end()) using the group_cursor_ to get the
// points within a group and lu_sim to draw the point realisations
int status = group_sequential_simulation(
*(neighborhood_.raw_ptr()), marginal,
progress_notifier.raw_ptr(),group_cursor_,
lu_sim );
// Back-transform the realisation contained in property prop from a
// Gaussian distribution to the target_cdf_
if( use_target_hist_ ) {
cdf_transform( prop->begin(), prop->end(),
marginal, target_cdf_ );
}
...
}
Function group_sequential_simulation
This function implements the sequential simulation on groups
instead of points. For each group along the random path (from
begin to end), it finds the points within that group with the object
Spectrum Series Volume 17 61
A BOUCHER

group_cursor, finds the neighboring data to that group of points

with the object neighbors and finally passes the unknown pixels
and the known values (both data and previously simulated
realisations) to the LU simulator.
int group_sequential_simulation( GeovalueIterator begin,
GeovalueIterator end, Neighborhood& neighbors,
const MarginalCdf& marginal, Notifier* notifier,
GroupCursor* group_cursor, GroupSimulator
group_simulator )
{
int bad = 0;
int loc = 0;
//Loop over all groups of nodes in the grid; the ordering has
//been already determined by the object path_. Geovalue is an
object that contains all the information about a point on a grid
such
//as its value and its location.
for(; begin != end; begin++, loc++) {
std::vector<Geovalue> group_gval;
group_cursor->get_group_gval(group_gval,*begin);
if( group_gval.empty() ) continue;
// Find the neighborhood data
neighbors.find_neighbors( * begin );
// Perform the simulation of the group of points by passing the
// unknown values (contained in group_gval.begin() to
// group_gval.end()), the known neighbouring data (contained in
// neighbors.begin() to neighbors.end()) and the marginal
// distribution (in this case a Gaussian with zero mean and unit
// variance) to group_simulator (in this case lu_sim).
// group_simulator returns zero when succesfull.
bad += group_simulator(group_gval.begin(),group_gval.end(),
neighbors.begin(),neighbors.end(),marginal);
// Send a signal to the progress bar and check if the user had
// cancelled the simulation in progress.
if( !notifier->notify() ) return -1;
}
return bad;
} FIG 2 - G-sgsim user interface.

User interface even for commercial products. By providing an interface to

The user interface, shown in Figure 2, is derived by adding a
widget to input the size of the group to be simulated. The QT
program Designer is used to modify the user interface – no actual
coding is involved for the interface changes in most cases.
Examples on a iron ore deposit
The Gsgsim plug-in described above is used to simulate an iron
deposit in Western Australia – see Boucher and Dimitrakopoulos
(2007) for more details on the deposit. The Gsgsim user interface
and the simulation parameters are shown in Figure 2. The drill
hole locations and Fe content are shown in Figure 3 and Fe
simulations generated with the plug-in with groups of 50 points
(5 × 5 × 2) are displayed in Figure 4. A total of 9 973 720 points
were simulated. Figure 5 shows a 3D rendering of the low iron
values, which correlate with high clay content.
Automating tasks with Python
It is often necessary to repeat tasks, eg to study the sensitivity of
an algorithm to different input parameters. Such sensitivity
analysis would be tedious if the user had to manually update the
parameters and re-run the algorithm. Embedded Python script in
SGeMS enables to automate such tasks. The Python language
(http://www.python.org/) is an interpretative language, ie a
program that does not require compilation to run. Python is
distributed under an open source license that makes it free to use,
Python, SGeMS allows users to automate repetitive tasks while
taking advantage of the rich variety of Python libraries. Python is
an object-oriented language with a long list of specialised
libraries to perform mathematical operations (eg numerical
integration and optimisation), image analysis, data visualisation
or even create graphical user-interfaces.
Some examples of common tasks performed with Python:
• Importing data – read data from any format and import them
into SGeMS.
• Exporting data – extract data from SGeMS and use Python to
export them into a desired format.
• Data processing – get data from SGeMS and modify data
values such as truncating high/low values, doing indicator or
logarithm transform, upscaling, computing local varying
means, etc.
• Sensitivity analysis – repeated calls of an algorithm with
different parameters.
• Calling external programs – make SGeMS interact with an
external program such as a mine schedule optimiser. This can
be a great time saver for workflows that need multiple calls
to programs.
The script must then do four steps:
1. export data from SGeMS (simulation, estimation, etc) in
the right format;

62 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 MINING GEOSTATISTICS WITH THE OPEN SOURCE SGEMS SOFTWARE AND THE GSTL LIBRARY

FIG 3 - Drill hole locations and Fe values.

FIG 4 - One realisation from the Gsgsim algorithm for an iron ore deposit.

2. write parameter file(s) for the external program;

3. run the external program(s) using Python; and
4. import, process and display the results in SGeMS.

Example – cross-validation
Cross-validation is a common tool for choosing the right
parameters for a given geostatistical algorithm. The fitness of a set
of parameters is measured by estimating locations that are known
but removed from the data set, and computing the errors between
these estimates and the true values. A zero error indicates un- FIG 5 - Three-dimensional rendering of the low simulated Fe
biasedness and smaller squared errors indicates a better accuracy. values indicating high clay content.
To perform cross-validation, the Fe drill hole data set presented
earlier is first partitioned into K non-overlapping set. Each finally automatically displayed on a graph as shown in Figure 6.
partition set is then successively removed from the data set and The ‘#’ character indicates comments.
estimated using the neighbouring data belonging to the K-1
remaining set. Each sampled location is estimated only once. The
performance of the algorithm is assessed by analysing the errors
kCrossValidation() function
between the true and the estimated values.
The Python script below performs cross-validation by looking # Cross-validation function
at errors in function of the variogram types (spherical, ex- # k : Numbers of partitions of the data
ponential or Gaussian) and the variogram range in the horizontal # grid : Name of the sgems grid
plane. The kCrossValidation function takes as input k, the # hd_prop : Name of the property containing the data
number of partitions for the data set, the name of the grid and # v_ranges: Vector containing the variogram ranges
property of the data set, the variogram ranges in x, y and z # v_type : String with the variogram type
direction and the type of variogram to consider for ordinary # Seed: Initialise the random number generator (Optional)
kriging. The main program gathers the parameter for each call of
the kCrossValidation function, considering variogram ranges def kCrossValidation(k, grid, hd_prop, v_ranges, v_type, seed =
varying from 20 to 500 per increment of 20. The process is 111) :
repeated for a spherical, exponential and Gaussian type of errors = 0.
variogram. The errors are collected in a vector, saved in a file and errors2 = 0.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 63
A BOUCHER

FIG 6 - Output from the Python script. Cross-validation errors (top) and squared errors (bottom) in function of the variogram range and type.

# Get the data from SGeMS kriging_values = scipy.array(sgems.get_property(grid,’__jk__’) )

data = scipy.array( sgems.get_property(grid, hd_prop) ) sgems.execute(‘DeleteObjectProperties
data_k = data.copy() dh::__jk__::__jk___krig_var’)
# Initialise the random number generator and randomise the # Compute the sum of errors and squared errors
indices of the data errors += scipy.sum(kriging_values[id] - data[id] )
scipy.random.seed( seed ) errors2 += scipy.sum( (kriging_values[id] -data[id])*
index = scipy.random.permutation( scipy.arange( len(data)) ) (kriging_values[id] - data[id]) )
# For each of the k partitions # Delete the temporary property and return an error vector
for i in range(k): sgems.execute(‘DeleteObjectProperties dh::__jk_hd__’)
# Get the indices for the k-th data partition (randomly selected) return [errors,errors2]
id = (index >= len(data)/k*(i)) == (index <
len(data)/k*(i+1)) Main program
# Set the k-th partition to uninformed and send the data to SGeMS
data_k[id] = -9966699 import sgems # functions to interact with SGeMS
sgems.set_property(grid,’__jk_hd__’, data_k.tolist()) import scipy # scientific computation module
# Replace back the uninformed data with the actual values import pickle # function to save complex data structures on file
data_k[id] = data[id] import pylab # plotting module
# Execute ordinary kriging with the specified variogram ranges # Create an empty dictionary to store the cross-validation errors
and variogram type errors = {}
sgems.execute(‘RunGeostatAlgorithm errors2 = {}
kriging::/GeostatParamUtils/XML::<parameters> # Vectors containing the variogram values to be tested
v_range = range(20,501,20)
<algorithm
v_ranges_str = [str(i) for i in v_range]
name=“kriging”/> <Grid_Name value=‘‘ ‘+grid+’ ” /> v_types = [‘Spherical’, ‘Exponential’, ‘Gaussian’]
<Property_Name value=“__jk__” /> <Kriging_Type # Loop all variogram types
type=“Ordinary Kriging (OK)” > <parameters /> for t in v_types :
</Kriging_Type>
errors[t] = []
<do_block_kriging value=“0”/> <Hard_Data
grid=‘‘ ‘+grid+’ ” errors2[t] = []
# Loop variogram ranges
property= “__jk_hd__” /> <Search_Ellipsoid
for r in v_ranges_str :
value=“500 500 10
# Call the cross-validation function with specific ranges (r) and
0 0 0” /> <Min_Conditioning_Data value=“0” />
variogram type (t)
<Max_Conditioning_Data value=“25” />
<Variogram results = kCrossValidation(50,’dh’,’FE_capped’,
[r,r,’15’] , t)
nugget=“0.2”structures_count=“1” > <structure_1
# Store the errors in the dictionaries
contribution=“0.8” type=‘‘ ‘+v_type+’ ” >
<ranges max=‘‘ ‘+v_ranges[0]+’ ’’ medium= ‘‘ errors[t].append(results[0])
‘+v_ranges[1]+’ ” errors2[t].append(results[1])
min=‘‘ ‘+v_ranges[2]+’ ” /> <angles x=“0” y=“0” # Save the results in a file
z=“0” /> ...Code omitted for brevity
</structure_1> </Variogram> </parameters>’) # Plot the errors and the squared errors in function of the range
# Extract and delete the kriging estimates from SGeMS ...Code omitted for brevity

64 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 MINING GEOSTATISTICS WITH THE OPEN SOURCE SGEMS SOFTWARE AND THE GSTL LIBRARY
This Python cross-validation code can easily be modified to
analyse the sensitivity to the nugget effect, anisotropy, search
neighbourhood or the relevance of using nested variogram
models.
CONCLUSIONS
SGeMS offers many opportunities for mining applications. The
extensive and open source geostatistical library provides a wide
range of geostatistical algorithms and a sound framework for
adding new algorithms. It offers opportunities to design algo-
rithms specific for a mining site – the parameters and the user
interface can be tailored for the needs of a particular application,
such as taking account of topographical trends, non-conventional
data transforms or even directly link to a company database. Some
parameters can be set to values chosen by experts for rapid on-site
use by general practitioners. The addition of Python scripts
provides a powerful way to streamline routine tasks and interface
with external programs. Finally, both academia and industry
would benefit from a large repository of plug-ins for mining
purposes where new algorithms are routinely contributed by users.
REFERENCES
Benndorf, J and Dimitrakopoulos, R, 2007. New efficient methods for
conditional simulation of large orebodies, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 103-110 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Boucher, A and Dimitrakopoulos, R, 2007. A new efficient joint
simulation framework and application in a multivariable deposit, in
Orebody Modelling and Strategic Mine Planning, second edition
(ed: R Dimitrakopoulos), pp 303-312 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Davis, M, 1987. Production of conditional simulations via the LU
decomposition of the covariance matrix, Mathematical Geology,
19:91-98.
Advances in Orebody Modelling and Strategic Mine Planning I
Deutsch, C and Journel, A, 1998. GSLIB Geostatistical Software Library
and User’s Guide, second edition (Oxford University Press).
Dimitrakopoulos, R and Luo, X, 2004. Generalised sequential Gaussian
simulation on group size v and screen-effect approximations for large
field simulations, Mathematical Geology, 36:567-591.
Journel, A G, 2007. Roadblocks to the evaluation of ore reserves the
simulation overpass and putting more geology into numerical
models of deposits, in Orebody Modelling and Strategic Mine
Planning, second edition (ed R Dimitrakopoulos), pp 29-33 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Osterholt, V and Dimitrakopoulos, R, 2007. Simulation of orebody
geology with multiple-point geostatistics – Application at Yandi
Channel iron ore deposit, WA and implications for resource
uncertainty, in Orebody Modelling and Strategic Mine
Planning, second edition (ed R Dimitrakopoulos), pp 51-59 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Remy, N, 2004a. S GeMs: The Stanford geostatistical modelling
software: A tool for new algorithms development in Geostatistics
Banff 2004 (eds: O Leuangthong and C V Deutsch), pp 865-872
(Springer: Dordrecht).
Remy, N, 2004b. Algorithmic and software methods for a better
integration of the geological information into numerical models,
PhD dissertation, Stanford University.
Remy, N, Boucher, A and Wu, J, 2009. Applied Geostatistics with
SGeMS: A User’s Guide (Cambridge University Press).
Remy, N, Schtuka, A, Levy, B and Caers, J, 2002. GsTL: The
geostatistical template library in C++, Computers and Geosciences,
28:971-979.
Strebelle, S, 2002. Conditional simulation of complex geological
structures using multiple-point statistics, Mathematical Geology,
34:1-21.
Wu, J, Boucher, A and Zhang, T, 2008. SGeMS code for pattern
simulation of continuous and categorical variables: FILTERSIM,
Computers and Geosciences, 34:1863-1876.
Spectrum Series Volume 17 65

Conditional Simulation Methods to Determine Optimum Drill
Hole Spacing
C Boyle1
ABSTRACT
Resource evaluation drilling is usually on a regular pattern; the pattern
spacing should be optimised to maximise profit from mining, accounting
for the cost of drilling and the value of additional information from
increased density of drilling, and also to reduce risk in mined ore tonnes
and grades to an acceptable level. Conditional simulation methods for
determining optimum drilling spacing are more powerful than traditional
methods, and simulation methods can take into account local variability
in grade. Mining profit functions can consider profit and density of
evaluation drilling so that profit can be maximised. A conditional
simulation method, here termed the simulation-estimation method,
consists of the following: making conditional simulations of the orebody
using the evaluation drilling; sampling these realisations to produce sets
of artificial drill holes; estimating resources using the artificial drill holes;
and comparing estimated resources with the realisation from which the
respective artificial drill holes were taken. The effectiveness of the
simulation-estimation method is demonstrated in a case study with the
exhaustive Walker Lake dataset in which predicted results, using a
sampled dataset, are similar to those from the exhaustive dataset. The
simulation-estimation method is applied to a mining case study of a
Western Australian bedded iron ore deposit and shows that estimation
errors are higher for the impurities SiO2, Al2O3, and P, rather than Fe. A
mining profit function suggests that the preferred resource evaluation drill
hole spacing to maximise mining profit is a spacing of 40 × 40 m, which
has a higher expected profit than spacings of 10 × 10 m, 20 × 20 m and 80 ×
80 m. Charts of drill hole spacing versus expected average error allow a
spacing to be selected that has the maximum allowable level of average
error. Conditional simulation methods can also be used for quantitative
resource classification that considers mining production rates and
selectivity, as well as for highlighting periods of high risk in mine plans
due to high uncertainty in ore grade.
INTRODUCTION
Resource evaluation drilling of ore deposits to allow estimation
of tonnes and grades is usually carried out on either a regular
square or rectangular grid pattern. Resource evaluation data is
needed for pit design, long-term scheduling, determining
blending requirements and stockpiling studies. As spacing of
drilling decreases, there is a rapid increase in cost for drilling,
sampling, assaying, and interpretation, while uncertainty of
grades and orebody location decrease. Optimum spacing can be
quantitatively evaluated using the cost of collecting extra
sampling versus economic benefits of this additional information
so that mining profit is maximised. Resource evaluation drilling
is usually not the final grade control method for allocation of ore
or waste categories. Final grade control is often done just prior to
mining, using additional sampling data collected from dedicated
reverse circulation (RC) percussion drilling or from blast holes.
Historically, ordinary kriging estimation variance was used to
determine accuracy of various drill hole spacings (David, 1988;
Journel and Huijbregts, 1978). Once enough evaluation drilling
data has been collected to produce a reasonable semivariogram,
various patterns of drill holes can be designed and dummy
ordinary kriging can be conducted based on the patterns of drill
holes. This can be done without actually drilling the holes, as
kriging variance only depends on the configuration of data values
and not on the magnitude of the values themselves. However, there
are a number of problems with the use of kriging variance to
determine the estimation error (Goovaerts, 1997; Schofield,
1. Email: cameron.boyle1@yahoo.co.id
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
2001; Chapman, 2003; Journel and Kyriakidis, 2004). Kriging
estimation variance is only the conditional variance in the case
of a Gaussian random function with strict stationarity, such that
simple kriging can be used as the estimation technique. The
distribution of estimation errors is likely to be skewed if the
variable estimated has a negatively or positively skewed
distribution, or where the point or block being estimated is in a
high or low grade area with respect to the deposit average. It is not
possible to sum kriging variances, as kriging variances of adjacent
blocks are spatially correlated and not independent. This makes it
impossible to compare kriging variance estimation errors for
various selective mining unit sizes.
Conditional simulation takes into account local grade
variability, such that areas with higher values can be modelled
with higher variances between different simulations
(Dimitrakopoulos, Godoy and Chou, 2010, in this volume).
Conditional simulation methods can determine appropriate drill
hole spacing in areas exhibiting the proportional effect (local
variance correlated with local mean) and heteroscedasticity
(fluctuating values of local mean and variance), rather than
homoscedasticity (constant mean and variance), as assumed in the
use of kriging variance.
METHODOLOGY
Relative absolute error
Simulation allows various sampling patterns and optimum drill
hole spacing to be assessed quantitatively by calculating various
measures like relative absolute error, amount of misclassification
of blocks in a resource model and the economic consequences of
estimation errors. With closer drill hole spacing there are more
conditioning data and the various simulations will be increasingly
alike. In the end case, if there could be complete sampling of the
deposit, then all simulations would be the same. One approach,
termed herein the simulation-estimation method, is documented
by Li et al (2007) and is as follows in the eight steps below:
1. Make a point simulation of the deposit at a fine scale
conditional to sampling and geological data. The size of the
blocks must be at least as fine as the closest drill hole
spacing to be tested later.
2. Sample simulations at a specified drill hole spacing to
generate a set of artificial drill holes.
3. Build a model of the deposit based on the set of artificial
drill holes, by means of the usual technique for resource
estimation of the deposit (often ordinary block kriging).
The model of the deposit is generated at the same block
size as for normal resource estimates.
4. Reblock the points in the fine initial simulation to the same
support as blocks in the resource model in step 3 above.
This reblocking is used to represent the same support size
used in reality, and must include a sufficiently large number
of nodes, so that the simulated models used represent the
actual variability of the blocks.
5. Calculate the relative absolute error of each block in the
estimated deposit developed in step 3 by comparing the
block grade to reality from step 4. This can be done on a
block by block basis, or a number of blocks can be
Spectrum Series Volume 17 67
C BOYLE
combined into volumes that represent mining over different
periods, such as months or years. Relative absolute error is
calculated by the formula:
E ij = Vsij - Ve j / Vsij (i = 1,..., n; j = 1,..., m) (1)
where:
Eij is the relative absolute error of block j estimated using
simulated deposit i
Vsij is the reblocked simulated value of block j
Vej is the estimated value of block j
n is the number of simulated deposits
m is the number of blocks
6. A relatively high number of simulations (eg 30 to 50) are
made at a fine scale, repeating step 1 above.
7. Repeat steps 2 to 5, using a variety of drill hole spacing
covering the range of possible drill hole spacing options.
8. Make charts of the average relative errors for different drill
hole spacings, both globally and for local areas. Maps can also
be produced showing relative errors for different regions.
Relative absolute error can be calculated for blocks of
selective mining unit size or on larger volumes after combining
adjacent blocks in a model. The larger volumes may represent
mining production for quarters or years, as determined from
long-term schedules of future mining. Grades of larger volumes
should be closer to true grades than grades of smaller volumes to
true grades, due to averaging of grades over the larger volumes.
Profit function
Drilling and sampling costs, as well as the potential benefits for
mine planning, can be calculated for various drill hole spacing
patterns. Conditional simulation allows distributions of profits for
resource model blocks to be determined and the spacing with
highest potential profit selected. Boucher, Dimitrakopoulos, and
Vargas-Guzman (2004) calculate profit per tonne mined for
various patterns of infill grade control drilling in a nickel cobalt
deposit and their approach is modified here for use in drilling of
an iron ore deposit at a resource evaluation stage, rather than at the
grade control stage.
Economic effects of block misclassification can be determined
using costs of sampling and drilling, and the profit from sale of
iron ore. US dollars are used throughout this study and all unit
costs and revenues are approximate for the general Australian
iron ore industry only, rather than for any particular mine or
company. The main costs are taken to be for drilling, logging,
sampling, and interpretation – as well as mining, crushing and
transportation from mine to port (by rail or truck). Costs of iron
ore shipping and later transport to steel mill are usually paid for
by the customer; these costs are not included in this study.
The cost of producing a block at location u in category k, is:
C kTotal (u) = C kDrilling + C kCrushing + C kTransportation (2)
where:
C kTotal is total cost
Drilling
C k is drilling, logging, sampling, and interpretation cost
C kMining is mining cost
Crushing
C is crushing cost
k The semivariogram model has a nugget of 25 000 (40 per cent),
Transportation
C k is transportation cost from mine to port
68 Spectrum Series Volume 17
The number of holes of a smaller drilling grid increases by a
multiplicative factor relative to the number of holes on a larger
grid size in the following relationship:
( )( )
fdrill hole = 1 + Xlarge / Xsmall . 1 + y large / y small / 4 (3)
where:
fdrill hole is a multiplicative factor giving the relative change in
number of drill holes with change in drill hole spacing
xlarge is the x axis spacing of the larger drill grid
xsmall is the x axis spacing of the smaller drill grid
ylarge is the y axis spacing of the larger drill grid
ysmall is the y axis spacing of the smaller drill grid
As an example, halving the grid spacing from 80 m to 40 m in
both x and y axes will increase the number of holes by a factor
fdrill hole of 9/4:
fdrill hole = (1 + 80/40)(1 + 80/40)/4 = 9/4
If 400 drill holes are needed to cover the extent of an orebody
with an 80 m grid, then 900 drill holes are needed to cover the
same area with a 40 m grid.
TESTING THE METHOD WITH AN EXHAUSTIVELY
KNOWN DEPOSIT
Data and statistics
The exhaustive Walker Lake elevation dataset consists of 78 000
measurements of a variable V from a regular 1 × 1 m grid that
extends 260 m east-west and 300 m north-south (Isaaks and
Srivastava, 1989). The exhaustive data set is sampled on a 20 m
square grid representing 13 lines (east-west) and 14 lines
(north-south) of drilling data from a mineral resource deposit at a
feasibility stage of evaluation (Figure 1). The sampled drilling
dataset is used with the simulation-estimation method to
determine optimum drill hole spacing. Six square drill hole
spacings are used:
1. 5 × 5 m,
2. 10 × 10 m,
3. 20 × 20 m,
4. 30 × 30 m,
5. 40 × 40 m, and
6. 50 × 50 m.
A spacing of 5 × 5 m refers to 5 mE × 5 mN and is for two
dimensions. Results are considered for 20 × 20 m squares,
representing blocks in a resource model. The 20 × 20 m areas are
reblocked into 60 × 60 m, 100 × 100 m and 260 × 300 m areas,
representing mining production for various time periods. The
Walker Lake study considers grade uncertainty for various areas,
as is required in the mining industry for stockpiling and blending
studies to manage product or crusher feed variability. The
proportions above or below a cut off grade are not considered for
the Walker Lake study.
The statistics and shape of the sample data set distribution
(Figure 2) closely resemble the distribution of the exhaustive
data set (Figure 3). The sample data set has a positively skewed
distribution with a moderately high coefficient of variation.
Directional semivariograms are calculated for V with a lag
distance of 20 m and directions aligned with the sample lines.
one spherical structure with a sill of 38 000 (60 per cent), and
Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL SIMULATION METHODS TO DETERMINE OPTIMUM DRILL HOLE SPACING
FIG 1 - Walker Lake sample locations.
FIG 2 - Histogram of variable V for samples at Walker Lake.
FIG 3 - Histogram of variable V for exhaustive data at Walker Lake.
ranges of 40 m east and 55 m north. Samples at 20 × 20 m,
representing evaluation drilling, are transformed to normal
scores by a Hermite polynomials method. Directional
semivariograms are calculated for normal scores of V and the
semivariogram model has a nugget of 0.2, one spherical
structure with a sill of 0.8, and ranges of 40 m east and 60 m
north.
Advances in Orebody Modelling and Strategic Mine Planning I
Simulation-estimation method
The simulation-estimation method uses the steps summarised
above. Fifty conditional simulations are made by the turning
bands method with 400 bands (Chiles and Delfiner, 1999; Boyle,
2007). The search neighbourhood is circular with radius 100 m, a
minimum of three and maximum of 16 samples, and no use of an
octant search (as drilling is not clustered). Simulations in normal
scores space are back transformed to data space. Point
simulations have no artificial striping, and honour means,
variances and semivariograms of sampling data. Point
simulations are reblocked to 20 × 20 m squares and 20 × 20 m
squares are reblocked to 60 × 60, 100 × 100 and 260 × 300 m
squares. Each simulation is sampled at six grid spacings (5 × 5,
10 × 10, 20 × 20, 30 × 30, 40 × 40 and 50 × 50 m) to produce six
sets of artificial drill holes. Ordinary block kriging estimates
20 × 20 m areas, for a total of 300 kriging estimates, that are
reblocked to four sizes (20 × 20, 60 × 60, 100 × 100 and
260 × 300 m). Relative absolute error for each area is calculated
by Equation 1 above and the simulation-estimation method has
50 reality datasets. Relative absolute errors for each of the four
area sizes and six spacings are shown in Figure 4.
True relative absolute error
Results from the simulation-estimation method are compared
with true results from the exhaustive data. True results are those
produced by ordinary kriging (as in a resource estimate) of
various spacings of drill holes extracted from the exhaustive
values, with the assumption of no sampling error. Estimates
made using true values are compared with the exhaustive data
reblocked to the same sizes (20 × 20, 60 × 60, 100 × 100 and
260 × 300 m). Averages of true relative absolute errors for each
of the four area sizes and six spacings are shown in Figure 4.
Results
Values of relative absolute error generally increase with
increasing drill hole spacing (Figure 4), though there are a few
exceptions due to whether specific samples are selected, for
example with 100 × 100 m areas between 40 m and 50 m drill
hole spacing. This is more pronounced for wider drill hole
spacings, as the influence of each sample is higher. The charted
lines for the simulation-estimation method plot within a
reasonable distance to lines of the true values, so the
simulation-estimation method generally performs acceptably in
FIG 4 - Summary of relative absolute errors of the simulation-
estimation method and relative absolute errors of true values for V
over four areas and using six drill hole spacings at Walker Lake.
Spectrum Series Volume 17 69
C BOYLE

this study. However, there is overestimation of average relative internal waste BIF that have separate interpreted wireframes, and
absolute error with closer spaced drilling patterns, for 10 m and samples of these BIF pods are not included in statistics and
20 m spacing in 20 × 20 m areas and for 10 m to 30 m spacing in simulation of the mineralised domain. Ideally, the full space of
260 × 300 m areas. At 30 to 50 m spacing in 20 × 20 m areas, the uncertainty of a mineral deposit is assessed by various geological
simulation-estimation method has similar results to true values. interpretations or domain simulations, for example as by
Drill hole spacing required for a defined relative absolute error Osterholt and Dimitrakopoulos (2007), but for this study only
depends on the area to be assessed. True values for an average deterministic wireframes are used and instead grade variability
relative absolute error of less than ten per cent for areas of within the ore domain is examined.
20 × 20 m require a 5 m square drilling spacing, whereas if the Fe is negatively skewed with a low coefficient of variation
areas are 260 × 300 m, then only 50 m drilling spacing is compared with Al2O3, SiO2 and P (Figures 5 to 8). Fe has strong
sufficient for a similar relative absolute error. This indicates a negative correlations with Al2O3 and with SiO2. Direct
requirement for early discussions between geologists and mining semivariograms of Al2O3, Fe, P and SiO2 are modelled in data
engineers with regards to planning drilling, expected mining space for use in estimation.
selectivity and production rates, so that drilling spacing is
optimised for evaluation. If production rates are high and
selectivity is low in a bulk mining operation, then early drilling
expenditure can be reduced by using a wider spacing, whereas
selective mining will require a tighter drill hole spacing if the
mine is to achieve expected production and cash flow.
In this Walker Lake study, involving one variable, one domain
and six possible drill hole spacings, the simulation-estimation
method uses 50 simulations, then 300 kriging estimations. There
is also the associated reblocking and calculations of relative
absolute error. Efficient software with a fast simulation algorithm
and scripting capabilities are prerequisites for handling the large
amount of processing in a drill hole spacing study by simulation
in a reasonable time. In summary, the simulation-estimation
method has performed reasonably in the Walker Lake study; the
same method is used in the following iron ore case study.

APPLICATION – THE WELD RANGE IRON ORE

DEPOSIT STUDY
FIG 5 - Histogram of Al2O3 in drill hole composites at Weld Range.
Data and statistics
The simulation-estimation method of determining drill hole
spacing for evaluation drilling is applied to the W14 orebody of
the Weld Range iron ore deposit (Neehoff et al, 1998). Weld
Range is located 65 km northwest of Cue, and 720 km
north-northeast of Perth in the Murchison area of Western
Australia. Mineralisation consists of a 4 km strike length of
massive goethite-haematite lenses, up to 80 m thick, dipping at
70° and located between a footwall of Archaean banded iron
formation (BIF) and hanging wall of dolerite. Canga (surface
re-cemented iron ore) is also present in addition to the massive
goethite-haematite lenses. Open file data provided by the
Western Australian Department of Industry and Resources from
the 1997 - 1998 drilling program at Weld Range are used for this
case study (Neehoff et al, 1998). This public domain drilling data
consists of 124 reverse circulation (RC) holes (WWRC01-
WRRC123) and 22 diamond holes (WR01-WR28), assayed for
Fe, SiO2, Al2O3, P, LOI, Mn, S, TiO2, Ca, MgO and K – with
density determination by downhole geophysical logging. Drilling
samples are composited to 2 m for equal support. These data are FIG 6 - Histogram of Fe in drill hole composites at Weld Range.
an example of what is available at a feasibility stage of resource
evaluation. Additional drilling as well as a feasibility study has
taken place at Weld Range since the 1997 - 1998 drilling Simulation-estimation method
program, but none of that company domain data has been used in
this study. The area modelled in this study is between 47 480 - Normal scores for Al2O3, Fe, P and SiO2 are created by a
50 520 mE (local grid). Hermite polynomials method. Direct and cross semivariograms
Optimum drill hole spacing depends on the most variable of of normal scores are modelled by a linear model of
the important chemical elements, rather than on less important or coregionalisation (Goovaerts, 1997) for use in cosimulation. A
more spatially continuous elements. For iron ore at Weld Range, spacing below of 10 × 10 × 2 m, for example, means 10 mE ×
the important four chemical variables have been identified as Fe, 10 mN × 2 mRL – three dimensions are considered in the Weld
SiO2, Al2O3 and P – only these are examined in this case study. Range study.
The majority of drilling samples have assays for all four
The eight steps for the Weld Range study are:
variables. Sectional drill hole interpretation defines a single main
mineralised domain of greater than 50 per cent Fe, hanging wall 1. Fifty point cosimulations of grid spacing 10 × 10 × 2 m are
dolerite and footwall BIF. The mineralised domain has pods of generated and checked.

70 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL SIMULATION METHODS TO DETERMINE OPTIMUM DRILL HOLE SPACING
FIG 7 - Histogram of P in drill hole composites at Weld Range.
FIG 8 - Histogram of SiO2 in drill hole composites at Weld Range.
2. Simulations are sampled at spacings of 10 × 10 × 2, 20 × 20
× 2, 40 × 40 × 2 and 80 × 80 × 2 m, representing 50 sets of
artificial drill holes.
3. Ordinary block kriging estimations are made using each
drill hole spacing for each sampled simulation, for a total
of 200 estimations. Ordinary block kriging is a common
estimation method for an iron ore resource model.
Estimations are for blocks of size 40 × 40 × 10 m.
4. The 50 initial simulations are reblocked to the same size as
for the blocks in the kriging estimations in step 3 above
(40 × 40 × 10 m). Reblocked simulations represent 50
representations taken to be reality.
5. Relative absolute error of each block in the estimated deposit
developed from a particular simulation in step 3 is calculated
by comparing estimated block grades to reality for that
simulation from step 4 (Equation 1). Errors are calculated
over the W14 orebody for production periods of quarters
(20 quarters of about 3 Mt) and years (five years of around
12 Mt). Tonnage calculations assume an in situ bulk density
of 3.2 kg/m³.
6. Divide the orebody into quarterly and yearly periods of
production as in a conceptual mine plan with 10 m mining
benches.
7. Charts of relative absolute errors are created for quarterly
and yearly production periods.
Advances in Orebody Modelling and Strategic Mine Planning I
8. Profits are calculated for each block when mine planning is
carried out using different evaluation drill hole spacings
and the spacing that maximises profit is determined, taking
into account the costs of drilling, mining, crushing and
transportation.
Results
Average relative absolute errors vary widely for Al2O3, Fe, P and
SiO2 for quarterly and yearly production periods (Figures 9 to
16). For each variable, the average errors of quarterly periods are
slightly higher than yearly periods for the same variable.
Averages of quarterly periods and yearly periods for the four drill
hole spacings are summarised in Figure 17. Fe errors are an order
of magnitude lower than for the other variables, due to the high
Fe concentration and lower coefficient of variation. Quarter 20
has a higher error than most other quarters, but even for quarter
20 on 80 × 80 × 2 m drill hole spacing, the average relative
absolute error of Fe is less than one per cent (Figure 10). Fe may
be less important than the contaminants in deciding drill hole
spacing due to the higher variability of contaminants. The
highest average relative absolute errors of contaminants are for
SiO2, followed by Al2O3 and then P. SiO2 relative absolute errors
average less than two per cent at 10 × 10 × 2 m spacing for both
quarterly and yearly averages, and increase monotonically to
greater than four per cent at 80 × 80 × 2 m spacing for both
quarterly and yearly periods (Figure 17). Various drill hole
spacings for different variables can be evaluated for their
respective average expected errors, and selection made of the
widest spacing with the required maximum allowable average
error. SiO2 may be the critical variable in determining drill hole
spacing due to its highest average relative absolute error and
importance as an impurity to steel mills. If average SiO2 relative
absolute error must be less than three per cent, then a spacing of
40 × 40 × 2 or 50 × 50 × 2 m is needed for both quarterly and
yearly periods of production (Figure 17).
Average relative absolute errors for quarters of each variable
differ considerably (Figures 9 to 12) due to mining areas with
different variability (heteroscedasticity). Due to correlation
between variables, periods with high Fe average errors often also
have high average errors for Al2O3, P and SiO2. Quarters 12 and
FIG 9 - Relative absolute errors for Al2O3 over quarterly periods
and estimated using four drill hole spacings at Weld Range.
Drill hole spacings are 10 × 10 × 2, 20 × 20 × 2, 40 × 40 × 2 and
80 × 80 × 2 m.
Spectrum Series Volume 17 71
C BOYLE

20 have high average errors predicted and are thus high risk,
whereas quarters 2 and 3 have low average errors predicted
(Figures 9 to 12). Average error charts for years mainly agree
with those for quarters – year 5 has higher than usual risk in
predicting ore grades (Figures 13 to 16). If expected grade
prediction errors are too high, then pre-emptive action can be
taken, for example by additional drilling or adding stockpiling
and blending capacity. Alternatively, multiple mining areas or
pits could be scheduled for periods of highest risk, allowing more
blending options and choice of mining blocks available to
overcome the higher expected variability. If mine plans rely on a
few mining areas of high average error, then the plans are high
risk. Expected financial penalties due to variability of impurities
in ore can be evaluated and an assessment of the value of
additional drilling can be made. Methods using conditional

FIG 12 - Relative absolute errors for SiO2 over quarterly periods

and estimated using four drill hole spacings at Weld Range.

FIG 10 - Relative absolute errors for Fe over quarterly periods and

estimated using four drill hole spacings at Weld Range.

FIG 13 - Relative absolute errors for Al2O3 over yearly periods and
estimated using four drill hole spacings at Weld Range. Drill hole
spacings are 10 × 10 × 2, 20 × 20 × 2, 40 × 40 × 2 and 80 × 80 × 2 m.

simulations can show areas of higher risk better than methods

FIG 11 - Relative absolute errors for P over quarterly periods and
estimated using four drill hole spacings at Weld Range.
using only drill hole spacing or kriging variances, as the latter
assume equal estimation error if sampling density is equal,
whereas errors are likely to be greater in areas with higher or
lower values of chemical variables than average.
Classification has been made into ore or waste for resource
model blocks, for each of 50 simulations (each taken to be
reality) and 50 sets of drill holes taken from sampling each
simulation at four drill hole spacings. Using each simulation as
reality, the proportion of misallocation of blocks can be
determined. In an iron ore deposit, grades of impurities can be as
important to steel mill customers as the grade of Fe. Three

72 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL SIMULATION METHODS TO DETERMINE OPTIMUM DRILL HOLE SPACING
FIG 14 - Relative absolute errors for Fe over yearly periods and
estimated using four drill hole spacings at Weld Range.
FIG 15 - Relative absolute errors for P over yearly periods and
estimated using four drill hole spacings at Weld Range.
chemical parameters are used for block classification in the Weld
Range study to define high grade ore for direct shipping to blast
furnaces:
1. Al2O3 <5 per cent,
2. Fe >57 per cent, and
3. SiO2 <10 per cent.
These high grade chemical parameters are generalised for this
study and are not from any particular mine. P has not been
considered in block classification, as the mean grade of the
samples is fairly low at 0.07 per cent and P levels should be able
to be controlled by blending during mining. Instead, Al2O3, Fe,
and SiO2 are more significant when deciding drill hole spacing.
Advances in Orebody Modelling and Strategic Mine Planning I
FIG 16 - Relative absolute errors for SiO2 over yearly periods and
estimated using four drill hole spacings at Weld Range.
FIG 17 - Summary of average relative absolute errors of the
simulation-estimation method over quarterly and yearly periods for
four drill hole spacings at Weld Range.
Proportions of ore and waste blocks are determined for
simulations (taken to be reality) and kriged estimates from
sets of drill holes are taken from each simulation. Block
misclassification decreases with decreasing drilling spacing
(Figure 18).
Prices paid for iron ore are based on metric tonne units (MTU)
of Fe metal in $US/MTU, either for fines or lump iron ore. This
study has been simplified to not consider lump and fines
percentages and uses a constant price of $US0.90/MTU
(EconStats, 2008), with no moisture adjustment. The revenue per
tonne of ore is calculated as:
Rore = Fe % x RMTU (4)
Spectrum Series Volume 17 73
C BOYLE

where: where:
Rore is the revenue per tonne of ore Pore-ore is profit for actual ore modelled as ore
Fe % is the Fe grade of ore
Pwaste-waste = -$5 - CDrilling (6)
RMTU is the revenue per metric tonne unit of Fe metal
where:
Pwaste-waste is profit for actual waste modelled as waste (always
negative)
Pore-waste = (Fe % × $0.90 - $5 - $1 - $3 - CDrilling) × 0.8 (7)
where:
Pore-waste is profit for actual ore modelled as waste
Pwaste-ore = -$5 - CDrilling (8)
where:
Pwaste-ore is profit for actual waste modelled as ore

Ptotal = ∑ i = 1 (Pore-ore + Pwaste-waste + Pore-waste + Pwaste-ore)

j
(9)

where:
Ptotal is total profit for all blocks 1 to j depending on the
category of each block j
FIG 18 - Weld Range proportion of ore and waste in mineralised Note that Pwaste-waste is assumed equal to Pwaste-ore,
ore domain. Legend format is <actual>_<modelled>, for example
in the legend ore_waste is for actual ore modelled as waste.
Paverage = Ptotal ÷ j (10)

In the case of 60 per cent Fe and $US0.90/MTU, the revenue where:

per tonne of ore is 60 per cent × $0.90/MTU = $54/tonne. Paverage is the average profit of all mined blocks 1 to j (ore and
Production costs per tonne in $US have been assumed to be: waste)
• mining $5, Paverage can be easier to use than Ptotal when selecting a drill hole
• crushing, spacing and is calculated for each drill hole spacing (Figure 19).
• screening $1, and
• trucking or railing to port $3.
Drilling, sampling, assaying, and interpretation costs, C kDrilling ,
are assumed to be $100 per metre for reverse circulation (RC)
drilling and on four possible drill hole spacings have costs per
tonne for ore and waste of 10 × 10 m at $1.084, 20 × 20 m at
$0.271, 40 × 40 m at $0.038 and 80 × 80 m at $0.017. There are
four categories, k, for mining blocks:
1. actual ore modelled as ore,
2. actual ore modelled as waste,
3. actual waste modelled as ore, and
4. actual waste modelled as waste.
It is assumed that later grade control will efficiently determine
the correct allocation of blocks (the information effect is not
taken into account in this study), mining blocks of 40 mE, 40 mN
and 10 mRL represent the selective mining unit, and that drilling
Drilling
costs C k are the same for ore and waste. Hence one cost
CDrilling suffices for all k (category). A discount factor of 0.8 is
applied to true ore estimated to be waste, with the assumption FIG 19 - Profit by drill hole spacing at Weld Range.
that most ore estimated to be waste lies inside the pit shell and
will be eventually located by grade control sampling, but that
timely mining of this ore will be delayed and mine planning will Even though a 10 × 10 m spacing will result in the most
be suboptimal. accurate estimation with the highest proportions of correctly
Profit calculations for each category, Pk, calculated with assigned blocks (Figure 18), the far higher cost of drilling for a
Equation 2 above are: 10 × 10 m spacing means that a 10 × 10 m spacing does not
maximise profit. For a spacing of 80 × 80 m, misallocation of ore
and waste blocks is too high to compensate for the lower cost of
Pore-ore = Fe % × $0.90 - $5 - $1 - $3 - CDrilling (5) drilling. Instead, a spacing of 40 × 40 m has the highest profit,

74 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL SIMULATION METHODS TO DETERMINE OPTIMUM DRILL HOLE SPACING
FIG 20 - Cumulative density function of relative absolute errors for 100 × 100 m areas at Walker Lake.
$US33.0/tonne, of four considered drill hole spacings and so
40 × 40 m is the preferred drill hole spacing at the resource
evaluation stage for the Weld Range iron ore deposit.
RESOURCE CLASSIFICATION
The JORC code for ore reserves and resources is prepared by the
Joint Ore Reserves Committee of the Australasian Institute of
Mining and Metallurgy, Australian Institute of Geoscientists and
Minerals Council of Australia and requests a statement of the
relative error or confidence in the resource estimate, if possible by
quantitative statistical or geostatistical procedures (Joint Ore
Reserves Committee, 1994). This can be done by conditional
simulation approaches as in the above studies of drill hole spacing.
Though there are currently no generally accepted quantitative
criteria for resource classification, Parker (2004) proposes that a
Measured resource classification is appropriate for volumes
of quarterly production where there is less than ±15 per cent
error with 90 per cent confidence, with confirmed continuity
of mineralisation. An Indicated resource classification is
appropriate for volumes of yearly production where there is less
than ±15 per cent error with 90 per cent confidence, with
assumed continuity of mineralisation. Error applies to both
tonnes and grades of ore. Conditional simulation can determine
expected error for different drill hole spacings and allow for a
competent person to decide resource classification. Applied to
the simulation-estimation method, these definitions for Measured
Advances in Orebody Modelling and Strategic Mine Planning I
and Indicated mean the relative absolute errors between kriged
grades or tonnages relative to simulated reality over the volumes
representing quarterly or yearly production are less than
15 per cent for 90 per cent of the volumes. Charting cumulative
distribution function (CDF) of block relative absolute errors
allows determination of error at a 90 per cent confidence level.
This approach is applied for the Walker Lake study, assuming a
quarterly production area is 100 × 100 m and a yearly production
area is 260 × 300 m. For 90 per cent of 100 × 100 m areas, the
relative absolute error is below 34 per cent for a 5 m drill hole
spacing and below 57 per cent for a 50 m drill hole spacing
(Figure 20), so no drill hole spacings considered allows a
Measured classification. For 90 per cent of 260 × 300 m areas,
relative absolute error is below eight per cent for a 5 m drill hole
spacing, below 12 per cent for a 30 m drill hole spacing and
below 22 per cent for a 50 m drill hole spacing (Figure 21). The
widest spacing with a relative absolute error below 15 per cent at
90 per cent confidence is the 30 m drill hole spacing, which
therefore allows an Indicated classification.
The Weld Range study considers relative absolute errors in
grade within a deterministic ore-BIF boundary. This study could
be extended to simulate ore-BIF boundaries (Osterholt and
Dimitrakopoulos, 2007), followed by grade simulation within the
simulated boundaries, to produce a more complete model of the
space of uncertainty. This would allow a quantitative
determination of tonnes and grade of ore that could be used for
resource classification at Weld Range.
Spectrum Series Volume 17 75
C BOYLE
FIG 21 - Cumulative density function of relative absolute errors for 260 × 300 m areas at Weld Range.
CONCLUSIONS
The simulation-estimation method performs reasonably for the
Walker Lake case study, as shown by comparison of average
relative errors of block grades calculated by ordinary kriging
using a sampled subset of simulations, when compared with
those of true kriging results expected from sampling of
exhaustive data. Errors predicted decrease as the size of the
production areas assessed increase – larger selective mining
units and higher production rates allow a wider evaluation drill
hole spacing to be used.
A profit function developed to determine optimum drill hole
spacing for resource evaluation and used in the Weld Range
iron ore case study, shows that maximum profit is achieved for
a 40 × 40 m drill hole spacing. Of the major chemical variables
for iron ore, average expected relative absolute error is greatest
to least for SiO2, Al2O3, P and Fe. For each variable, average
errors for quarterly production periods are slightly greater than
for yearly production periods for the same variable. Charts of
drill hole spacing versus average relative absolute error show a
suitable spacing for the required level of estimation error.
Resource classification can be determined quantitatively for
expected production rates and selectivity. Mine plans can be
examined using conditional simulation methods to show
periods of high risk and action, like extra drilling or plan
changes, taken if risk is unacceptable.
76 Spectrum Series Volume 17
REFERENCES
Boucher, A, Dimitrakopoulos, R and Vargas-Guzman, A, 2005. Joint
simulations, optimal drillhole spacing and the role of the stockpile, in
Geostatistics Banff 2004, Proceedings Seventh International
Geostatistics Congress, Canada, pp 35-44 (Springer: Netherlands).
Boyle, C, 2007. Comparison of multivariate conditional simulation
methods at the Yandicoogina iron ore deposit, in Proceedings Iron
Ore 2007, pp 57-66 (The Australasian Institute of Mining and
Metallurgy: Melbourne).
Chapman, E, 2003. An investigation using conditional simulation to assess
the relationship between the level of confidence in the estimated
resource and the drilling grid at the Warramboo pisolite iron ore
deposit, Robe River, Masters of Science in Mining Geology thesis
(unpublished), Camborne School of Mines, University of Exeter,
Cornwall, United Kingdom.
Chiles, J P and Delfiner P, 1999. Geostatistics, Modelling Spatial
Uncertainty, 695 p (John Wiley and Sons: New York).
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
Reserve Estimation, 217 p (Elsevier: The Netherlands).
Dimitrakopoulos, R, Godoy, M and Chou, C L, 2010. Resource/reserve
classification with integrated geometric and local grade variability
measures, in Advances in Orebody Modelling and Strategic Mine
Planning I (ed: R Dimitrakopoulos), pp 215-222 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
EconStats, 2008. Iron ore prices and producers [online], EconStats.
Available from: <http://www.econstats.com/rt_ironore.htm>
[Accessed: 10 November 2008].
Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL SIMULATION METHODS TO DETERMINE OPTIMUM DRILL HOLE SPACING
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation, 483 p
(Oxford University Press: New York, USA).
Isaaks, E and Srivastava, R, 1989. An Introduction to Applied
Geostatistics, 561 p (Oxford University Press: New York).
Joint Ore Reserves Committee, 2004. The JORC Code, Australasian
Code for Reporting of Exploration Results, Mineral Resources and
Ore Reserves. Available from: <http://www.jorc.org> [Accessed 10
November 2008].
Journel, A and Huijbregts, C, 1978. Mining Geostatistics, 600 p
(Academic Press: London).
Journel, A and Kyriakidis, P, 2004. Evaluation of Mineral Reserves a
Simulation Approach, 232 p (Oxford University Press: New York).
Li, S, Dimitrakopoulos, R, Scott, J and Dunn, D, 2007. Quantification of
geological uncertainty and risk using stochastic simulation and
applications in the coal mining industry, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 253-260 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Advances in Orebody Modelling and Strategic Mine Planning I
Neehoff, B, Miller, A, Dumlao, F and Llorca, J, 1998. An Feng Kingstream
Steel Weld Range Project TR70/3902 and E20/176, Exploration report
for the period September 1997 - November 1998, Department of
Industry and Resources Western Australia, open file report A56957.
Osterholt, V and Dimitrakopoulos, R, 2007. Simulation of orebody
geology with multiple-point geostatistics – Application at Yandi
Channel iron ore deposit, WA and implications for resource
uncertainty, in Orebody Modelling and Strategic Mine Planning,
second edition (ed: R Dimitrakopoulos), pp 51-59 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Parker, H, 2004. Resource and Reserve Audits, Geostatistical Association
of Australia talk, Perth, Western Australia, 10 May.
Schofield, N, 2001. Determining optimal drilling densities for near mine
resources, in Mineral Resources and Ore Reserve Estimation – The
AusIMM Guide to Good Practice, (ed: A Edwards), pp 293-298 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Spectrum Series Volume 17 77

Transformation Methods for Multivariate Geostatistical
Simulation – Minimum/Maximum Autocorrelation Factors and
Alternating Columns Diagonal Centres
E M Bandarian1, U A Mueller2, J Fereira3 and S Richardson4
ABSTRACT
To speed up multivariate geostatistical simulation it is common to
transform the set of attributes into spatially uncorrelated factors that can
be simulated independently. The main method in recent years has been
minimum/maximum autocorrelation factors, either based on the
coefficient matrices of a two structure linear model of coregionalisation
(LMC) or on a pair of experimental covariance matrices. In both cases
there is an underlying assumption that the covariance structure of the data
set can be adequately modelled using a two structure LMC. We consider
an extension that removes the restriction imposed by this assumption by
using the experimental matrices for a larger set of lags. The method relies
on an iterative algorithm that approximately diagonalises a set of
symmetric matrices, and is referred to as the Alternating Columns-
Diagonal Centres method. We use the Jura data set to evaluate the extent
to which factors obtained from each method are spatially decorrelated
and to assess the effect of the transformation method on the simulated
attributes.
INTRODUCTION
The simulation of one or more attributes in a multivariate data set
is enhanced through the exploitation of the relationships between
them. However, multivariate modelling and simulation can be
time consuming and computationally inefficient. Some of the
shortcomings of multivariate methods can be avoided by
transforming the spatially correlated set of attributes into a set of
spatially uncorrelated factors. These factors can then be modelled
and simulated independently using univariate geostatistical
methods. The oldest method for obtaining uncorrelated factors is
principal component analysis (Wackernagel, 2003). While this
technique is easy to implement, the factors are uncorrelated for all
distances other than zero only in the case of intrinsic correlation.
In recent years the method of minimum/maximum autocorrelation
factors (MAF) has received much attention as the factors are
derived so that they are approximately spatially uncorrelated for
all distances (Desbarats and Dimitrakopoulos, 2000; Boucher
and Dimitrakopoulos, 2007). This method assumes that the data
to be modelled are multivariate normal and so in most cases the
first step is to normalise the data. While the method of direct
minimum/maximum autocorrelation factors (DMAF) (Bandarian,
Bloom and Mueller, 2008) removes the requirement of normality
of the data and so simplifies the MAF procedure, one still requires
the assumption of a two structure linear model of coregionalisation
(2SLMC) in order to determine the transformation coefficients.
Extensions of MAF to more than two structures are not possible
(Vargas-Guzman and Dimitrakopoulos, 2003), unless the model
1. Mathematics Teacher, Cyril Jackson Senior Campus, 53 Reid Street,
Bassendean WA 6054. Email: ebandari@cyriljackson.wa.edu.au
2. Associate Professor, School of Engineering, Edith Cowan University,
270 Joondalup Drive, Joondalup WA 6027.
Email: u.mueller@ecu.edu.au
3. Honours Student, School of Engineering, Edith Cowan University,
270 Joondalup Drive, Joondalup WA 6027.
Email: jferrer0@ecu.edu.au
4. Lecturer, School of Engineering, Edith Cowan University,
270 Joondalup Drive, Joondalup WA 6027.
Email: s.richardson@ecu.edu.au
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
can be described as a nugget model together with an intrinsic
spatial component (Tran, Murphy and Glacken, 2006). The use of
a 2SLMC to derive the transformation matrix is largely of
theoretical interest and while there are versions of MAF that
work directly with the model (Bandarian, 2008), in most
implementations it is common to work with experimental matrices
directly as proposed by Desbarats and Dimitrakopoulos (2000),
thus cutting down on the effort spent in finding a reasonably well
fitting 2SLMC. As a trade-off however time needs to be invested
in choosing the lag spacing at which to calculate the experimental
semi-variogram matrix. Whichever approach is chosen, be it
theoretical or experimental, the factors calculated from the input
data are only approximately spatially orthogonal.
Because of the restrictions in the MAF-method, a more general
approach such as joint approximate diagonalisation (JAD) can be
considered. JAD consists of finding a transformation matrix that
will approximately diagonalise a set of matrices. These techniques
have been developed for image-filtering and blind noise separation
(see for example Yeredor (2002) and references therein). JAD
algorithms for approximate diagonalisation of symmetric matrices
can be broadly split into two categories: those that derive an
orthogonal diagonalising matrix like that proposed by Manton
(2005) and those that weaken the orthogonality constraint to non-
singularity of the diagonaliser. The algorithm we will consider is
of the latter type and is referred to as the Alternating Columns
Diagonal Centres (ACDC) method (Yeredor, 2000, 2002). For
geostatistical applications the set to be jointly approximately
diagonalised consists of experimental semi- variogram matrices.
The transformation matrix derived from this method will
approximately spatially decorrelate the given attributes, in the
sense that the cross correlations between distinct attributes are
approximately equal to zero.
In the standard implementation, MAF is applied to normal
scores data and the resultant factors are subsequently simulated.
There is no theoretical requirement to apply MAF to standard
normal data (Bandarian, 2008) and we will be taking the
approach to calculating the required transformation matrices
from standardised raw data instead. The reason for this approach
is partly the observation that the application of MAF to normal
scores can result in transformed data that are no longer normally
distributed, thus resulting in the need for a further normal scores
transform if a Gaussian simulation algorithm is to be used.
In what follows, we present a brief review of the MAF and
ACDC methods and illustrate them using a subset of the Jura
data set (Goovaerts, 1997). The transformations are calculated
from the correlation matrix and semi-variogram matrices of
standardised data. The factors resulting from each of the methods
are transformed to normal scores in order to implement sequential
Gaussian simulation (SGS). The normal scores simulations are
back-transformed to factors, then to the original data space. The
resultant simulations are assessed for reproduction of the spatial
and statistical characteristics of the sample.
PROBLEM DATA
A four variable subset of the Jura data set (Goovaerts, 1997)
consisting of Cd, Co, Cr and Ni is used to illustrate each
transformation method and the resultant simulations. Location
maps (see Figure 1) for the sample data show no obvious trends
Spectrum Series Volume 17 79
E M BANDARIAN et al

FIG 1 - Location maps for cadmium, cobalt, chromium and nickel. The colour scales are based on the deciles of the respective sample data.

TABLE 1
Summary statistics for the input data, for each attribute there are 259 values.

Variable Mean Variance Skewness Minimum LQ Median UQ Maximum

Cd 1.31 0.84 1.50 0.14 0.63 1.07 1.72 5.13
Co 9.30 12.79 -0.18 1.55 6.52 9.76 12.00 17.72
Cr 35.07 120.07 0.29 8.72 27.44 34.84 42.32 67.60
Ni 19.73 67.78 0.16 4.20 13.72 20.56 25.44 53.20

in the spatial distribution of the data. The sample data have quite TABLE 2
disparate means and variances (see Table 1) and of the four
Correlation matrix for sample data.
metals, only Chromium is approximately normally distributed
(Bandarian, 2008). The sample correlation coefficients indicate Variable Cd Co Cr Ni
varying strengths of linear relationships, with those between
Cd 1.00 0.25 0.61 0.49
Cd-Cr, Co-Ni and Cr-Ni having the stronger correlations (Table 2).
For stability of implementation the data are first standardised by Co 0.25 1.00 0.45 0.75
subtracting the sample mean from each value then dividing by Cr 0.61 0.45 1.00 0.69
the sample standard deviation.
Ni 0.49 0.75 0.69 1.00

MULTIVARIATE TRANSFORMATION METHODS

AND SIMULATION
T
Let Z(u)=[Z1(u), Z2(u),...ZK(u)] be a vector of K second-order
stationary random functions over a study region A, let
z(u)=[z1(u ), z2(u ),...zK(u )]T, α=1,...n, be the corresponding
isotopic sample vectors and let B denote the correlation matrix of
the random function.
The general procedure for applying transformation methods
for multivariate simulation is:
step 1 – transform the data to approximately decorrelated
factors,
step 2 – independently model and simulate factors, and
step 3 – back transform simulated factors to simulated attributes.
Step 1 includes the calculation of the transformation matrix
from the data, in the case of MAF this requires in the first
instance a decision what lag spacing to choose for the calculation
of the experimental semi-variogram matrix, for ACDC a decision
needs to be made as to how many semi-variogram matrices
are to be included in the computation. In addition in step 1
an assessment needs to be made whether or not the resultant
factors are sufficiently well decorrelated. Assuming the spatial

80 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 TRANSFORMATION METHODS FOR MULTIVARIATE GEOSTATISTICAL SIMULATION

decorrelation is satisfactory, then step 2 is only concerned with experimental semi-variogram Γ$ (h1 ) matrix calculated at lag
the univariate simulation of the factors. Either a direct simulation spacing h1 can be used to approximate Γ(h) at h1. In this case the
algorithm, such as the sequential simulation proposed by Soares MAF transformation matrix is derived from the generalised
(2001) or a standard Gaussian algorithm, such as Sequential eigenvalue problem reformulated in terms of the correlation
Gaussian Simulation can be used. In the latter case a Gaussian matrix B and the experimental semi-variogram matrix Γ$ (h1 ) at
anamorphosis will need to be carried out to compute normal lag h1 where h1 < a (a is the maximum range of the
scores for the factors. For both methods, Step 3 consists merely semi-variogram). Regardless of the suitability of the 2SLMC the
of the backtransformation of the simulated factors via matrix transformation matrix X simultaneously diagonalises Γ$ (h1 ) and
multiplication. The resultant realisations are still standardised B yielding factors which are orthogonal at lag spacings zero and
and so shifting and scaling to restore the appropriate mean and h1 only.
variance respectively concludes the workflow. A crucial assumption in the construction of the MAF factors is
the requirement that the covariance structure of the underlying
Minimum/maximum autocorrelation factors random function is fully characterised by a 2SLMC. This
assumption is restrictive in practice and extensions to more
For MAF it is assumed that the semi-variogram function, (h), of general models have been discussed in the literature (see for
the multivariate random function Z(u) can be modelled by a two example the discussion in Vargas-Guzman and Dimitrakopoulos
structure linear model of coregionalisation (h)=B1g1(h)+B2g2(h), (2003). However, in general, for three covariance structures or
where the symmetric coregionalisation matrices B1 and B2 more there exists a matrix X that diagonalises the variogram only
contain the sills of the permissible semi-variogram models g1(h) when all but (at most) one of the coefficient matrices in the LMC
and g2(h), and B=B1+B2 is the correlation matrix. Since B1 and are proportional to one another (Tran, Murphy and Glacken,
B2 are symmetric (therefore diagonalisable) and B2 is positive 2006; Bandarian, 2008).
definite (has positive eigenvalues), then B1 and B2=B-B1 may be
diagonalised simultaneously by congruence (that is, there exists a
non-singular matrix X, not necessarily orthogonal, such that Alternating Columns Diagonal Centres
S=XTAX) and the diagonalising matrix is the solution to the
symmetric definite generalised eigenvalue problem (Datta, 1995): The Alternating Columns Diagonal Centres (ACDC) method
(Yeredor, 2002, 2004) iteratively determines a matrix X that
B1X=(B-B1)X diagonalises a set of J symmetric K × matrices {B1,B2,...,BJ}
The diagonal matrix in the equation above is the matrix of by minimising:
generalised eigenvalues and the matrix X is the matrix of J
Ψ(XD1,D2,...,DJ) = ∑ w j B j − XD j XT
2
generalised eigenvectors. ,
F
j=1
The equation may be rewritten as:
B1X=BX 1
where ⋅ F denotes the Frobenius norm, wj > 0, j = 1,..., J are
(optional) weights and {D1,D2,...,DJ} is a set of diagonal
where: Λ1=Λ(I+Λ) -1
matrices. The algorithm consists of alternating phases, the AC
The generalised eigenvalue problem may be converted into a phase, where the objective function is minimised with respect to
standard eigenvalue problem with a symmetric matrix on the left
the columns of X while the set {D1,D2,...,DJ} is kept fixed and
hand side of the equation using the Cholesky factorisation of the
matrix B where B=LLT and L is a non-singular lower triangular the DC phase, where the objective is minimised with respect to
matrix. If we put G=L-1B1(LT)-1 and Y=LTX, then GY=YΛ1. {D1,D2,...,DJ} while X is kept fixed.
The matrix G is symmetric by construction and so Given {B1,B2,...,BJ} the algorithm is initialised by specifying
orthogonally diagonalisable. The matrix X obeys either an initial set of diagonal matrices {D1,D2,...,DJ} or an
XTBX=YTL-1B(LT)-1Y=YTL-1LLT(LT)-1Y=YTY=I and the estimate of the diagonalising matrix X. The initial phase in the
column vectors in X are B-orthonormal, that is XTBX=I. iteration is dependent on the specification made: if an initial
Furthermore, XTB1X=YTL-1B1(LT)-1Y=YTGY=Λ1. guess for the diagonalising matrix is made, then the algorithm
starts with a DC phase, otherwise the starting point is an AC
Thus the non-singular matrix X, whose columns are the phase. In the absence of either specification the diagonalising
generalised eigenvectors xk,k=1,...,K, simultaneously diagonalises matrix is set to the identity matrix and the algorithm is initialised
the matrices B1 and B by congruence (Bandarian and Mueller, with a DC phase. Thus the standard iterations are as follows:
2008).
given {B1,B2,...,BJ},
The MAF factors are derived by putting FMAF(u)=XTZ(u). $ = I.
put X
Since the matrix X is B-orthogonal the factor variances are
equal to one and the transformation ensures that the factors F(u)
are uncorrelated for all h ≥ 0 because: DC phase
Γ F(h) = X Γ(h)X
T
1. $ TX⊗X
Set G = [X $ TX]-1
T T
= X B1Xg1(h)+X (B-B1)Xg2(h) 2. For j = 1,..J
= Λ1g1(h)+(I-Λ1)g2(h) $ = diag(G diag(X
Set D $ TB X)).
$
j j
Hence Γ F(h) is diagonal for all h.
The MAF transformation thus diagonalises the semi-variogram AC phase
model of the attributes exactly. However, the associated
experimental semi-variograms of the MAF-factors are only For k = 1,...K
J
⎡ K
⎤
$ [ j] ⎢ $ T ⎥
approximately diagonalised. 1. Set P = ∑ w j λ k B j − ∑ λ n xnx n
[k ]
$ $
⎢ ⎥
Where the theoretical LMC is not known the experimental j=1
⎢⎣
n =1
⎥⎦
semi-variogram matrices Γ$ (⋅) are used to calculate the MAF
n ≠k

transformation coefficients. Assuming that the semi-variogram 2. Find the largest eigenvalue of P and an associated
function of Z(u) is fully characterised by a 2SLMC, an eigenvector of unit norm.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 81
E M BANDARIAN et al

a μ Perfect spatial decorrelation occurs when ζ(h) = r(h) = 0 and

3. If μ < 0 set x$ k = 0 otherwise put x$ k = k.
κ h) = 1 for all lag vectors h. A set of factor semi-variogram
∑w ( )
J
λ$ [kj ]
2
j matrices may be considered to be nearly in diagonal form if κ(h)
j=1 ≥ 0.9 for all h lags (Xie, Myers and Long, 1995).
The algorithm alternates between the two phases and halts Global measures for the spatial decorrelation are the averages
once a prespecified tolerance has been reached. The algorithm is ζ, τ and K calculated over J lag spacings:
applied to a sequence of semi-variogram matrices, calculated at
1 J 1 J 1 J
increasing distance from the origin. There are no requirements
other than real symmetry of the matrices. If X denotes the
ζ= ∑
J j=1
ζ(h j ), τ = ∑ τ(h j ) and K = ∑ k(h j )
J j=1 J j=1
matrix that approximately diagonalises the given semi-variogram
matrices then the ACDC factors are derived putting Simulations and reproduction of sample
FACDC (u ) = X−1 Z(u ). Thus for both methods the factors are characteristics
derived using a linear transformation, however neither of the
matrices is orthogonal, as would be the case for PCA. The For each attribute, the suite of realisations is assessed for
transformations are both local transformations, in that they act on reproduction of the attribute (target) histogram, semi-variogram,
the attributes at each location, but the matrix coefficients are mean, variance and correlation coefficients. Qualitative analysis
constant across the entire study region. Provided the factor scores consists of visual inspection of the histogram and semi- variogram
are sufficiently spatially orthogonal the factors may be swarms of the realisations and boxplots of the realisation means,
independently modelled and simulated. In this paper, a standard variances and correlation coefficients about the target statistic.
Gaussian algorithm will be used for the factors. As a first step Quantitative measures for the reproduction of the target
the factors are transformed to normal scores (Gaussian statistics are the semi-variogram and histogram mean square
anamorphosis), then they are independently modelled and deviation (VMSD and HMSD respectively) which are calculated
simulated using Sequential Gaussian Simulation (100 for each realisation. The HMSD is given by:
realisations). The simulated attribute scores are retrieved via
1 P
Gaussian anamorphosis from normal scores to factor scores, then
from factor scores to standardised scores using the inverse MAF
HMSD = ∑ ( z p − z1p )2
P p= 1
and ACDC transformation matrices. Finally the sample means
and standard deviations are reincorporated. where:
zp is the pth percentile of the target cumulative distribution
Performance measures function (cdf)
zlp is the pth percentile of the lth simulated cdf
Spatial decorrelation
P denotes the total number of percentiles calculated
The quality of the spatial decorrelation of the factors will be The VMSD is given by:
assessed both graphically and numerically. For a graphical
assessment the cross semi-variogram for each factor pair will be 1 J
graphed in order to detect any remaining spatial correlation.
VMSD = ∑ ( γ$ j − γ$ lj )2
J j=1
Quantitative measures (Tercan, 1999) used to assess spatial
decorrelation are the absolute deviation from diagonality ζ(h), where γ$ j and γ$ lj are the target and simulated experimental
the relative deviation from diagonality τ(h) and the spatial semi-variogram values respectively calculated at the jth lag
diagonalisation efficiency K(h). spacing and J is the total number of lags.
The absolute deviation from diagonality, ζ(h) at lag h is
defined to be the sum of squares of the off-diagonal elements of RESULTS AND ANALYSIS
the factor experimental semi-variogam matrix at lag h:
Transformations
ζ(h) = ∑ k = 1 ∑ j ≠ k ( γ$ (h; k, j)) , h > 0
K K 2
For ACDC, the set of target matrices consists of thirteen
F
experimental semi-variogram matrices calculated at a lag spacing
where γ$ F (., k, j) denotes the experimental cross-semi-variogram of 0.2 km with a lag tolerance of 50 per cent. The weight vector
for the factors Fj and Fκ . was set to be w = [10 10 1 ... 1]T. The transformation matrix is:
The function τ(h) compares the absolute sum of off-diagonal ⎡ 0.841 − 0.048 − 0.289 − 0152
. ⎤
~ ⎢ 0.049
elements of the factor experimental semi-variogram matrix Γ F(h) 0.782 0.525 − 0.811⎥
with the sum of the absolute values of the diagonal elements X−ACDC
1
=⎢ ⎥
calculated at each lag h: ⎢−0.382 − 1.045 1.247 0.457⎥
⎢ 0132 − 0.966 1.939 ⎥⎦
⎣ . 0121
.
∑ ∑
K K
k=1 j≠k
γ$ F (h; k, j)
τ(h) = ,h > 0 The MAF transformation matrix, shown is obtained using the
∑ γ$
K
F (h; k, k) sample correlation matrix B and an experimental semi-variogram
matrix Γ$ (h)1 , calculated at h1 = 0.220 km:
k=2

Finally, the function K(h) compares the sum of squares of the ⎡−1.082 0.215 − 0.036 0.288⎤
off diagonal elements of the factor experimental semi-variogram ⎢−0.663
matrix Γ$ F(h) at lag h to the sum of squares of the off diagonal − 0.934 1.005 0.695 ⎥
XTMAF =⎢ ⎥
elements of the sample experimental semi-variogram matrix ⎢−0114
. 1176
. 1.062 − 1.493 ⎥
Γ$ z(h): ⎢−0152 − 0.305 − 0.955 ⎥⎦
⎣ . 0.485

∑ ∑
K K
k=1 j≠k
γ$ F (h; k, j))2 This spacing was chosen from a set of separation distances as it
κ(h) = 1 ,h > 0 provided the best overall decorrelation for MAF (Bandarian,
∑ ∑
K K
k=1 j≠k
( γ$ z (h; k, j)) 2
2008).

82 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 TRANSFORMATION METHODS FOR MULTIVARIATE GEOSTATISTICAL SIMULATION

FIG 2 - Experimental cross semi-variograms of factors.

The cross semi-variograms for each factor set are displayed in of these measures for all lags (0.037 km to 2.407 km inclusive)
Figure 2. For ACDC the spatial decorrelation is excellent with and for the lower lags (0.037 km to 1.199 km inclusive) are
the majority of cross semi-variogram values satisfying shown in Table 4.
−015
. ≤ γ ik (h) ≤ 015
. . For this method there are only three
instances where the factor pairs still show some correlation. For TABLE 4
MAF, as expected, the spatial decorrelation is perfect at
Average (h), (h) and (h) for all lags and lower lags (0.037 km to
h1 = 0.220 km, while for other lag spacings the semi-variogram 1.199 km inclusive).
values are typically −0.2 ≤ γ ik (h) ≤ 0.2 (with the exception of
factors 2 and 4 at h1 = 1199
. km). These results are summarised in
All Lower
Table 3.
Plots of the measures of spatial decorrelation (ζ(h), (h) and ξ(h) τ(h) κ(h) ξ(h) τ(h) κ(h)
(h)) are displayed in Figure 3 and the corresponding averages ACDC 0.065 0.157 0.975 0.055 0.166 0.970
MAF 0.076 0.202 0.968 0.071 0.210 0.957

TABLE 3
Summary of factor pairs with remaining spatial correlation. The plots of ζ(h), (h) and (h) reflect the excellent spatial
decorrelation achieved by ACDC, which outperforms MAF for
Method Factor pairs Lag spacing γik
(km)
all lag spacings other than 0.220 km. The average diagon-
alisation efficiency for ACDC is 0.975 for all lags and 0.97 for
AC-DCAC-DC A3-A4 0.797 0.227 the lower lags, which is slightly higher than that achieved by
1.199 0.228 MAF where the values are 0.97 for all lags and 0.96 for the lower
2.407 −0.300 lags.
MAF D1-D3 0.595 −0.182 The factor scores from each method are transformed to normal
scores using a Gaussian anamorphosis with 50 Hermite
1.405 0.168
polynomials to approximate the factor histograms, independently
1.795 0.160 modelled then simulated using SGS. In order to avoid spurious
D2-D3 1.012 −0.163 correlations resulting from the use of the same random paths for
D2-D4 0.797 −0.182 each factor, the random number seeds were changed, so that
corresponding factors for the two methods used the same seed,
1.199 −0.221 but within the simulation of the factors the seeds were distinct.
D3-D4 2.196 0.178 The simulated attribute scores are retrieved via backtransformation
from normal scores to factor scores, then from factor scores to
standardised scores using the inverse ACDC and MAF
transformation matrices:
⎡ 1.258 0.416 0.277 0.207 ⎤
⎢−0189
. 1.024 − 0131
. 0.4445⎥
XACDC =⎢ ⎥
⎢ 0.215 0.862 0.660 0.222 ⎥
⎢ 0.033 0.585 ⎥⎦
⎣ 0.337 0.318

⎡−0.926 − 0.247 − 0.094 − 0.271⎤

⎢ 0.030 0.009 − 0.471 − 0.882 ⎥
(X )
−1
T
=⎢ ⎥
⎢−0.382 − 0.592 − 0.691 − 0164
MAF
. ⎥
⎢−0121 − 0.621 − 0.229 − 0.740 ⎥⎦
⎣ .
Finally the sample means and standard deviations are
FIG 3 - Plots of spatial decorrelation (ζ(h): n, (h): ∗ and (h): p). reincorporated.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 83
E M BANDARIAN et al

FIG 4 - Mosaic maps of realisation 25 for Cd, Co, Cr and Ni.

Simulations
Mosaic maps for one of the realisations (picked at random)
generated for each metal are shown in Figure 4. The mosaic
maps are broadly similar. The spatial variability of the sample
data has been reproduced in the simulation.

Reproduction of the target statistics

The realisation cdf swarms overlaid with the corresponding
sample cdf for each attribute and transformation method indicate
that the attribute distributions have generally been reproduced
(see Figure 5). The most notable deviation from the target cdf
occurs for Co. For both methods the Co target cdf is poorly
reproduced on the interval from approximately 4 ppm to 6 ppm
with the realisation swarms being consistently lower than the
target cdf. This is more pronounced for ACDC than for MAF.
Boxplots of the HMSDs for each set of attribute realisations
(Figure 6) reflect the overall reproduction of the target cdfs. For
Cd and Co the HMSDs from MAF are slightly lower than those
for ACDC, for Ni and Cr, the situation is reversed. For Cd the
HMSDs are higher for ACDC than those for MAF, with the
former having 75 per cent of values less than 0.209 while the
latter has 75 per cent of values less than 0.160. The distribution
of HMSDs for Co is similar for both methods although those for
MAF are generally slightly lower than for ACDC. For Cr the
HMSDs for ACDC are typically lower than those for MAF with
the former having 75 per cent of values less than 2.007 while the
latter has 75 per cent of values less than 2.238. Similarly for Ni
where ACDC has 75 per cent of values less than 1.385 while FIG 5 - Histogram swarms.
MAF has 75 per cent of values greater than 1.507.
Figure 7 displays the boxplots of the realisation summary means are within ±5 per cent of the target for both methods. The
statistics. For both methods and all attributes the target means MSD of the means for each attribute (Table 5) are generally
have been reproduced with the majority of realisation means lower for MAF, with the exception of Co where ACDC is lower.
being within ±5 per cent of the target. For Cd, ACDC and MAF Reproduction of the target variances is similar for both
80 per cent and 84 per cent of realisation means respectively lie methods. In general the realisation variances are lower than the
within ±5 per cent of the target. For Co this increases to 98 per corresponding targets, however the majority of realisation
cent and 97 per cent respectively while for Ni the percentages are variances are within ±15 per cent of the target for both methods.
96 per cent and 97 per cent respectively. For Cr the realisation For ACDC Cd, Co, Cr and Ni have 84 per cent, 78 per cent

84 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 TRANSFORMATION METHODS FOR MULTIVARIATE GEOSTATISTICAL SIMULATION

Reproduction of the target correlation coefficients (see Figure 8)

FIG 6 - Boxplots of histogram mean square deviation.
100 per cent and 62 per cent of realisation variances within
±15 per cent of the target respectively. For MAF these
percentages are 86 per cent, 84 per cent 98 per cent and 58 per
cent respectively. The MSD of the variances for each attribute
(Table 5) are generally lower for MAF, with the exception of Ni
where ACDC is lower.
TABLE 5
Mean square deviation of realisation means and variances.
between attributes is similar for both methods. With the
exception of Cd-Co the targets have been reproduced with the
majority of realisation correlations being within ±10 per cent of
the target. For Cd-Cr, Co-Ni and Cr-Ni in excess of 96 per cent
of realisation correlations are within ±10 per cent of the target
for both methods. For Cd-Ni the percentages are 78 per cent and
75 per cent for ACDC and MAF respectively, while for Co-Cr
they are 84 per cent and 69 per cent respectively. For Cd-Co the
target correlation coefficient has not been reproduced by either
method, although to a lesser extent for ACDC than for MAF.
ACDC yields 27 per cent of realisation correlations within
±10 per cent of the target while for MAF this percentage is only
19 per cent. The MSD of the correlations for each attribute pair
(Table 6) are similar for both methods, with ACDC being slightly
lower for Cd-Co, Cd-Ni and Co-Cr while for the remaining
attribute pairs MAF is slightly lower.
TABLE 6
Mean square deviation of correlation coefficients.
MSD Cd-Co Cd-Cr Cd-Ni Co-Cr Co-Ni Cr-Ni
Correlations AC-DC 0.065 0.029 0.040 0.035 0.040 0.026
MAF 0.072 0.028 0.041 0.043 0.038 0.024

MSD Cd Co Cr Ni
Reproduction of the target variograms
Means AC-DC 0.051 0.198 0.673 0.527 Experimental semi-variograms for the MAF and ACDC
MAF 0.045 0.223 0.515 0.462 simulations in Figure 9 indicate that both approaches have resulted
in adequate reproduction of the experimental sample semi-
Variances AC-DC 0.091 1.635 8.198 9.827
variograms. The variograms swarms are similar for both
MAF 0.088 1.485 7.439 9.974 transformation methods for each attribute and attribute pair. In

FIG 7 - Boxplots of realisation means and variances.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 85
E M BANDARIAN et al

general the direct semi-variograms have been reproduced with

FIG 8 - Boxplots of realisation correlation coefficients.
FIG 9 - Experimental semi-variogram swarms (grey lines) and
corresponding targets (•).
the overall shapes, ranges, nuggets and sills of the simulations
being similar to those of the target variograms. The most notable
exception to this is the nugget for Cd (both transformations)
where the target is somewhat lower than those of the simulations.
For the cross semi-variogram swarms, the simulations
obtained from both transformation methods are again similar. In
general the target cross semi-variograms have been reproduced
with respect to the overall shapes, nuggets and sills (except
Cd-Co where the target sill is underestimated in both cases).
With the exception of Cd-Cr the target ranges have been
overestimated. For Co-Ni, the cross semi-variogram swarm from
ACDC reproduces the target cross semi-variogram slightly better
than the swarm resulting from MAF, in particular for lag
distances between 1 km and 1.5 km. However, for both methods
the cross semi-variograms have lower sills than the experimental
data. This is also featured in the correlation coefficients for this
attribute pair, for Co and Ni the correlation at lag 0 is lower than
that of the target.
The overall reproduction of the direct and cross semi-
variograms is reflected by the low values of the VMSDs for all
attributes and attribute pairs for both transformation methods
(Table 7 and Table 8). The majority of VMSDs are less than
0.15, with the exception of those for Co-Ni where the majority of
VMSDs are less than 0.2. The distribution of VMSD values is
generally similar for each transformation method, with results for
ACDC slightly better than for MAF.
For both ACDC and MAF, the transforms are constructed
from omnidirectional semi-variogram matrices. Of the data
under consideration Ni and Co exhibit anisotropy with direction
of greatest continuity approximately E-W for Ni and SW-NE
for Co. A comparison of the average experimental semi-
variograms of the simulations and the standardised Co data
indicates that both methods capture the anisotropy and that
there is almost no difference in the average directional
variograms calculated from the realisations (Figure 11).
Similarly for Ni, the average experimental semi-variograms
reproduce the E-W anisotropy (Figure 12). For Cr and Cd, the
average semi-variograms show the realisations to be isotropic.
Thus at the univariate level, the spatial features of the raw data
are broadly reproduced, even though the spatial decorrelation
methods were based on omnidirectional semi-variogram
matrices.
Similarly, an inspection of the directional cross-semivario-
grams for Co and Ni and the corresponding realisations reinforces
the observations from the omnidirectional cross-semivariogram
swams. There is good coincidence between the average
experimental cross variograms for both sets of realisations and the
corresponding data cross-variograms for shorter separation
distances, while for distances between 1 km and 1.5 km the sills in
directions SW-NE and E-W do not reach the data sills, in direction
SE-NW the averages fit well.

TABLE 7
Mean square deviation of semi-variogram values.

ACDC MAF
Cd Co Cr Ni Cd Co Cr Ni
min 0.098 0.043 0.049 0.064 0.096 0.04 0.061 0.071
lq 0.104 0.055 0.06 0.097 0.101 0.049 0.072 0.096
med 0.108 0.062 0.066 0.109 0.104 0.056 0.076 0.106
uq 0.111 0.07 0.073 0.122 0.108 0.064 0.08 0.117
max 0.118 0.09 0.095 0.15 0.115 0.087 0.091 0.149
mean 0.108 0.063 0.068 0.109 0.105 0.057 0.076 0.106
variance 0 0 0 0 0 0 0 0

86 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 TRANSFORMATION METHODS FOR MULTIVARIATE GEOSTATISTICAL SIMULATION

TABLE 8
Mean square deviation of cross-semi-variogram values.

Cd-Co Cd-Co Cd-Cr Cd-Cr Cd-Ni Cd-Ni Co-Cr Co-Cr Co-Ni Co-Ni Cr-Ni Cr-Ni
(ACDC) (MAF) (ACDC) (MAF) (ACDC) (MAF) (ACDC) (MAF) (ACDC) (MAF) (ACDC) (MAF)
min 0.039 0.043 0.054 0.062 0.033 0.039 0.044 0.041 0.041 0.062 0.041 0.045
lq 0.069 0.074 0.067 0.072 0.079 0.08 0.068 0.075 0.121 0.112 0.077 0.091
med 0.083 0.088 0.08 0.081 0.101 0.102 0.085 0.093 0.141 0.137 0.098 0.108
uq 0.111 0.108 0.094 0.099 0.128 0.12 0.101 0.111 0.171 0.164 0.12 0.126
max 0.146 0.175 0.137 0.158 0.182 0.205 0.17 0.184 0.226 0.238 0.179 0.216
mean 0.088 0.093 0.083 0.086 0.104 0.103 0.088 0.097 0.143 0.138 0.098 0.109
variance 0.001 0.001 0 0 0.001 0.001 0.001 0.001 0.002 0.001 0.001 0.001

FIG 10 - Experimental cross semi-variogram swarms (grey lines) and corresponding targets (•).

CONCLUSION availability in commercial software, even though the need to use a

In this study, we have presented two approaches to the simulation
of a multivariate data set which rely on successful spatial
decorrelation of the raw data, standardised to unit variance and
mean zero, prior to simulation. The approach differs from the
standard approach in that the data were not transformed to normal
scores prior to the application of the decorrelation algorithms.
However, in contrast to the implementation in Bandarian (2008),
rather than using a direct sequential simulation algorithm, the
factors were simulated using sequential Gaussian simulation.
Thus, it was necessary to the convert each factor to normal scores
and back-transform after completion of the simulations, the latter
operation being automatic in sequential Gaussian simulation. The
choice of a standard algorithm was largely prompted by its ready
non-linear transformation may well be regarded as a disadvantage.
The data used for our study were a subset of the Jura data set
and the variables chosen are linearly correlated, an aspect which
is important as it is known that MAF does not cope well with
non-linear relationships between the data (see for example
Rondon and Tran, 2008). The four variables in our subset
provided a mix of isotropic and anisotropic data, making the use
of an algorithm such as sequential Gaussian co-simulation
unattractive as a parsimonious model of coregionalisation would
have had to be chosen for the simulation. For both methods the
anisotropy in the data is broadly reproduced, both at the
univariate and the bivariate level, even though omnidirectional
semi-variogram matrices were used for the derivation of the
factor transformations.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 87
E M BANDARIAN et al

FIG 11 - Directional experimental semi-variograms for standardised cobalt compared to mean directional semi-variograms for Alternating
Columns Diagonal Centres simulations and minimum/maximum autocorrelation factors simulations.

FIG 12 - Directional experimental semi-variograms for standardised nickel compared to mean directional semi-variograms for Alternating
Columns Diagonal Centres simulations and minimum/maximum autocorrelation factors simulations.

88 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 TRANSFORMATION METHODS FOR MULTIVARIATE GEOSTATISTICAL SIMULATION
The results show that both decorrelation methods are viable
for the simulation of the Jura data set. The results indicate that
ACDC performs at least as well as MAF for this set. Both
decorrelation methods considered use approximate diagonalisation
via a congruence transformation. The main difference lies in the
assumption that a 2SLMC describes the LMC for the data in the
case of MAF, which is not needed for ACDC. The ACDC
method thus gives the user greater flexibility and is as easy to use
as the MAF decorrelation method. Given the performance for the
Jura data, the ACDC method for decorrelation has the potential
to be used successfully with data sets whose LMC is not fully
characterised by two structures.
REFERENCES
Bandarian, E M, 2008. Linear transformation methods for multivariate
geostatistical simulation, PhD thesis (unpublished), Edith Cowan
University, Perth.
Bandarian, E M, Bloom, L M and Mueller, U A, 2008. Direct
minimum/maximum autocorrelation factors for multivariate
simulation, Computers and Geosciences, 34:190-200.
Bandarian, E M and Mueller, U A, 2008. Reformulation of MAF as a
generalised eigenvalue problem, in Proceeding of the Eighth
International Geostatistics Congress (eds: J Ortiz and X Emery),
pp 1173-1178.
Boucher, A and Dimitrakopoulos, R, 2007. A new efficient joint
simulation framework and application in a multivariable deposit,
in Orebody Modelling and Strategic Mine Planning, second edition
(ed: R Dimitrakopoulos), pp 345-354 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Datta, B N, 1995. Numerical Linear Algebra and Applications, (Brookes/
Cole Publishing Company: Pacific Grove).
Desbarats, J A and Dimitrakopoulos, R, 2000. Geostatistical simulation
of regionalized pore-size distributions using min/max autocorrelation
factors, Mathematical Geology, 32(8):919-942.
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
483 p (Oxford University Press: New York).
Advances in Orebody Modelling and Strategic Mine Planning I
Manton, J H, 2005. A centroid (Karcher mean) approach to the joint
approximate diagonalisation problem: The real symmetric case,
Digital Signal Processing,16(5):465-478.
Rondon, O and Tran, T T, 2008. Multivariate simulation using min/max
autocorrelation factors: Practical aspect and case studies in the
mining industry, in Proceedings of the Eighth International
Geostatistics Congress (eds: J Ortiz and X Emery), pp 269-278.
Soares, A, 2001. Direct sequential simulation and cosimulation,
Mathematical Geology, 31(2):155-173.
Tercan, A E, 1999. Importance of orthogonalization algorithm in
modelling conditional distributions by orthogonal transformed
indicator methods, Mathematical Geology, 31(2):155-174.
Tran, T T, Murphy, M and Glacken, I, 2006. Semivariogram structures
used in multivariate conditional simulation via minimum/maximum
autocorrelation factors, in Proceedings XI International Congress,
IAMG, Liege.
Vargas-Guzman, J A and Dimitrakopoulos, R, 2003. Computational
properties of min/max autocorrelation factors, Computers and
Geosciences, 29:715-723.
Wackernagel, H, 2003. Multivariate Geostatistics, third revised edition,
387p (Springer-Verlag: Berlin).
Yeredor, A, 2000. Approximate joint diagonalization using
non-orthogonal matrices, in Proceedings International Workshop on
Independent Component Analysis and Blind Source Separation
(ICA2000), pp 33-38.
Yeredor, A, 2002. Non orthogonal joint diagonalization in the least square
sense with application in blind source separation, IEEE Signal
Processing, 50(7):645-648.
Yeredor, A, 2004. ACDC: Approximate joint diagonalisation (in the
direct Least-Squares sense) of a set of Hermitian matrices, using the
iterative ACDC algorithm [MATLAB code]. Available from:
<www.eng.tau.ac.il/~arie/>. [Accessed: 3 May 2009].
Xie, T, Myers, D E and Long, A E, 1995. Fitting matrix-valued variogram
models by simultaneous diagonalization (Part II: Application),
Mathematical Geology, 27:877-888.
Spectrum Series Volume 17 89

Conditional Wavelet Based Simulation of Non-Stationary
Geologic Fields Using Geophysical and Model Analogue
Information
E Gloaguen1 and R Dimitrakopoulos2
ABSTRACT
Non-stationarity and scale dependencies are critical topics when modelling
spatial phenomena of complex geological patterns. A two-dimensional
conditional simulation based on wavelet decomposition of geophysical
and model analogue data is proposed for simulating non-stationary
geological patterns. The proposed method requires that geophysical data
are available on the entire area under study, and that we have an idea of
the small scale patterns of the geology (model analogue). Then, the
method utilises the wavelet transform of the geophysical data and of the
model analogue to decompose both of them into wavelet coefficients at
different scales, and then quantifies their spatial dependencies. Using
cokriging, the approximation coefficients of the variable under study are
inferred from the approximation coefficients of the geophysical data.
Joint simulation of the wavelet coefficients of the variable under study is
employed together with available soft conditioning data. The conditionally
cosimulated wavelet coefficients are back transformed generating
realisation of the attribute under study. The proposed method is applied
on a synthetic channel example. Two models are generated, one represent-
ing the ‘real geological model’ and the other one, is the model analogue.
The model analogue represents the same stratigraphy (channels, number of
geological units) but differs by the spatial location and by the proportion of
the channels. The synthetic geophysical data consists in full coverage of
the model under study. The model analogue is transformed into a physical
property model in order to compute the forward modelling of the
geophysical property. The results show that wavelet simulation of the
geological field shows good agreement with the synthetic one both in
terms of amplitudes and shape of the channels.
INTRODUCTION
The modelling and assessment of spatial uncertainty of the
attributes of geological phenomena are frequently based on the
stochastic simulation of random fields conditional to available
data. Conventional simulation methods (Chilés and Delfiner,
1999; Goovaerts, 1997) represent the geological complexity in
terms of second order statistics at a single scale, thus limiting
their ability to model more complex features (Guardiano and
Srivastava, 1993). Here scale refers to the support size on which
a given pattern is defined. During the past years, substantial
efforts have been made to develop new techniques that may take
into account the so-called high-order spatial statistics (Mustapha
and Dimitrakopoulos, 2010, in this volume) including multiple-
point statistics (Guardiano and Srivastava, 1993; Strebelle,
2002), computer graphic approach (Zhang, Switzer and Journel
2006; Arpat and Caers, 2004), Markov random field sequential
simulation (Daly, 2004) and directional metropolis for non-linear
likelihood posterior updating (Tjelmeland and Eidsvik, 2004). As
the two last methods are based on data, the multiple point
geostatistics approach (Guardiano and Srivastava, 1993) uses a
training image instead of a two-point variogram to account for
geological information. The training image describes the
geometrical facies patterns believed to be present in the
subsurface and reflect a prior geological/structural concept. This
1. Institut National de la Recherche Scientifique, 490 rue de la
Couronne, Québec QC G1K 9A9, Canada.
Email: Erwan_Gloaguen@ete.inrs.ca
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montréal QC H3A 2A7, Canada.
Email: Roussos.dimitrakopoulos@mcgill.ca
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
new approach allows the combination of the easy-conditioning of
pixel-based algorithms with the ability to reproduce shapes of
object-based techniques, without relying on excessive cpu
demand (Strebelle, 2002). Recently, research on developing
simulation methods with multiple point statistics uses
convolution of the training image with spatial filters yielding to
filter scores that permits the clustering of similar spatial patterns
in the filter score space (Zhang, Switzer and Journel, 2006).
The local multiple point statistics is captured from the training
image, and reproduced by simulating patterns. Hence, the complex
multiple-scale geological structures can be reconstructed in the
simulation grid, conditional to a variety of subsurface data such
as well data and seismic survey. A major limitation of existing
high-order statistics simulations is that they do not consider or
carry consistent relations between patterns at different scales, the
topic of ongoing research.
The relation between patterns at different scales can be assessed
through discrete wavelet transform or DWT (Portilla and
Simoncelli, 2000). The properties of DWT coefficients fully
characterise an image in terms of scale dependencies, features and
statistics (Mallat, 1989). DWT decomposes an image into wavelet
coefficients that are regrouped in directional sub-bands related to a
given scale (or level) of an image and allows both multiple scale
analysis and geostatistical modelling (Flandrin, 1992). The
dependencies of wavelet coefficients at different scales, namely,
high-order dependencies of wavelet coefficients, are studied for
statistical image modelling with applications to texture analysis
and texture synthesis (Crouse, Nowak and Baraniuk, 1998). These
techniques use hidden Markov models (HMM) and Gaussian
mixture model (GMM). Gaussian mixture model maximised the
model likelihood between the data and the mixture model to
capture the marginal statistics. HMM attempts to take into account
the interscale correlation that can be seen as a multidimensional
GMM that applies to tree-structured Markov chains across scales
(Choi and Baraniuk, 2001). The limitations of these techniques
include their inability to allow direct conditioning to data, only the
correlation between scales is taken into account and the likelihood
modelisation is an computer intensive iterative process. Using a
different wavelet based approach Tran, Mueller and Bloom (2002)
have shown that unconditional Haar wavelet simulations can give
results equivalent to the classical sequential Gaussian simulations
in the case of multi-Gaussian data.
This paper proposes a method for direct conditional simulation
in two dimensions, based on the DWT coefficient of a indirect
geophysical measurement and geological analogue. First, the
DWT of the geological and corresponding geophysical analogues
is used to capture the high-order dependencies between features at
different scales between the geological and geophysical data. Also,
the high-order dependencies of the geological features is explored
through covariances and cross-covariances modelling of the DWT
coefficients. The covariance captures the spatial relation between
directional wavelet sub-bands representing a given resolution and
the cross-covariances capture the frequency relation between the
directional wavelet sub-bands at different scales. Subsequently,
realisation of the DWT coefficients are joint simulated using a
sequential Gaussian cosimulation approach (Verly, 1994). Finally,
the DWT coefficients are backtransformed in the space of the
attribute of interest using the inverse DWT. A key aspect of the
method is that it allows conditioning to any measured data (hard
Spectrum Series Volume 17 91
E GLOAGUEN and R DIMITRAKOPOULOS

conditioning) or conditioning only at a given scale based on

indirect information (soft conditioning).

TWO-DIMENSIONAL DISCRETE WAVELET

TRANSFORM
In this section the two-dimensional DWT is presented and
important scaling properties of wavelet coefficients is revisited in
the context of geostatistical simulation. Then, conditioning to
hard and soft data from analogue petro-physical scenario is
presented.

Two-dimensional wavelet transform revisited

By construction, wavelet transform provides both spatial and
frequency-domain information about a regularly sampled grid
such as an image (Mallat, 1989). Wavelet analysis procedures
involve the decomposition of an image in terms of a set of
translated and dilated wavelet functions ΨB (where B = {H, V, D})
and scaling functions (φLL) that form an orthonormal basis of the
Hilbert space L2(2) of square integrable functions (Fan and Xia,
2003).
Similarly to the two-dimensional Fourier Transform, the
columns of an image are analysed with the scaling functions
(φLL) and bandpass filters ΨB, and the rows of each of these
resultant images are analysed again with each of (φLL) and ΨB.
This filtering step results in four sub-bands for each scale of
decomposition j, where j=1,...,J; the approximation characterising
the omnidirectional low frequency variations and three details
representing the local directional variations in three directions,
horizontal, vertical and diagonal (H, V, D).
For a given scale decomposition j=1,..,J, DWT of an image
x(s,t) of size N N, where s and t are the Cartesian coordinate
axis, is given by (Fan and Xia, 2003; Crouse, Nowak and
Baraniuk, 1998):
Nj −1 J Nj −1
x(s,t)= (s,t)+ Z Bj , k , i ψBj , k , i ( s , t ) (1)
k, i = 0 B = {H ,V , D} j = 1 k , i = 0
where:
−j
φ LL
J , k, i = 2
2 φ( 2 − j s − k, 2 − j t − i ) and
−j
N
ψ BJ , k , i = 2 2
ψ B ( 2 − j s − k, 2 − j t − i ) with Nj = . Property 3: In theory, the DWT should decompose an image
2j
a J, k, i = x ( s, t )φ j 0 , k , i dsdt are the approximation coefficients and
Z Bj , k , i = x ( s, t )ψ Bj , k , i dsdt are the (k,i)th wavelet coefficients of
sub-band B in scale j.
In the following paragraphs, the important properties of wavelet
decomposition are defined in the context of two-dimensional
simulation with scale dependency preservation. To make the
notation lighter, the wavelet coefficients or sub-band coefficients
at a given level j, Z Bj , k , i are noted Hj, Vj and Dj. Each level
represents a given scale of the image. Figure 1 shows the two-
dimensional decomposition tree of an image x(s,t) for J = 3
decomposition levels. The link between the different scales is
represented by the arrows, pointing toward the scale j from the
generating scale j-1. As an example, in Figure 1, each decom-
position scale produced one approximation coefficient map (a3)
and three directional detail coefficient maps (H3, V3, D3).
Practical aspects of the wavelet coefficients
Property 1: DWT provides with a set of directional and
frequency dependent coefficients that can be represented on a
FIG 1 - Three discrete wavelet transform scales. All the subgrids
represent the same area. Arrows point from the parent (scale j-1)
toward the child (scale j). Hj, Vj and Dj are the directional wavelet
detail coefficients at scale j and a3 is the approximation at scale
J=3.
tree structure (Kumar and Foufoula-Georgiou, 1997). This
property is important as DWT decompose the image in sets of
oriented coefficient at different scale allowing to characterise
possible geometric anisotropies at different scales. Note that each
sub-band represents the same area as the original training image
x(s,t).
Property 2: DWT allows taking into account the opposite
relation between resolution and support size. As a reminder, this
principle tells that the more precise the reslolution is, the smaller
the support size should be, and vice versa. For example, the
small grids on the diagonal coefficients (Figure 1) represent the
same area but the resolution of each child is four times less than
the resolution of its parent. The explicit modelling of the
cross-covariances between features at different scales is
particularly interesting for down-scaling purposes.
with uncorrelated wavelet coefficients representing the most
important features of an image, but Crouse, Nowak and Baraniuk
(1998) demonstrated that there are still considerable dependencies
between scale, as observed from the characteristics of wavelet
coefficient distribution (Fan and Xia, 2003). This proves that there
is still scale dependency in the wavelet coefficient. The
dependencies between scales are classified in two groups:
intrascale and interscale as discussed in the next sections.
By definition (Equation 1), at a given scale, the three directional
decompositions are collocated and represent the directional
component of an image at a given frequency (for example H3, V3,
D3 and a3 in Figure 1). The cross-variance between the sub-bands
belonging to a same scale characterises the spatial link between
the directional features of an image at this scale. Equation 1 also
represents the link between the smallest structures of an image and
the largest structures being consistent with the trade off relation
between resolution and support size. The cross-covariance
between two sub-bands belonging to two different scales
characterises the spectro-spatial information shared between the
two sub-bands (for example H3 and V1 in Figure 1). The wavelet
coefficients at different scales are nested. Consequently, starting
from the coarsest level to a point in the finest level is similar to
zooming without conflict between scales. This property is
particularly important for scaling purpose (Mallat, 1989).

92 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL WAVELET BASED SIMULATION OF NON-STATIONARY GEOLOGIC FIELDS

Intra- and interscale cross-variances of this particular spatial each sub-band, the cross-variograms between sub-bands belonging
decomposition is a direct link to high-order statistics (Portilla to both same scale or not and any known wavelet coefficient data.
and Simoncelli, 2000) as each point at the coarsest scale is In the following paragraph, the sequential Gaussian cosimulation
related to each of its generating points belonging to the finer is presented in the two-dimensional wavelet coefficient frame-
scales. Once the cross-covariances of the inter- and intrascales work. The detailed theory of Sequential Gaussian CoSimulation
are modelled, the spatial statistics are then of the order of the (SGCS) can be found in Verly (1994).
number of level times three, the number of sub-bands.
For mathematical simplicity, we present only the estimation of
Inverse discrete wavelet transform (IDWT) the wavelet coefficients belonging to a same scale J=1 at location
x0. There are four covariables to estimate, Za1,ZH1, ZV1 and ZD1.
Inverse discrete wavelet transforms are used to back transform
conditionally cosimulated wavelet coefficients to realisations of Assuming the variables are joint multinormally distributed and
the attribute under study. IDWT is the back transformation from considering the already measured or previously simulated matrix
the wavelet space to the original space. Inverse discrete wavelet Z = [ Znd
a 1 H1 D1 V 1 T
Znd Znd Znd ] of size nd known values by four variables,
transforms can be performed by inverting the orthogonal matrix
that transforms the original coefficients into the coefficients in the conditional distribution of Zxa01 ZxH01 ZxD01 ZVx 01 at x0 is defined by
−1
the wavelet basis. In practice, the IDWT is computed starting its conditional mean m( Zxa01nd , H 1, D 1
) = C x 0 nd C ndnd Z and its
from the coarsest scale and combining the approximation and the a 1 , H 1 , D 1 ,V 1
conditional variance C ( Zx 0 nd −1
) = C x 0 x 0 − C x 0 nd C ndnd C ndx 0 .
details into the approximation at the next finer scale, and so on.
Where DWT involves filtering and downsampling, the IDWT where:
consists of upsampling and filtering. Downsampling by p is
defined as taking every pth sample, starting with sample 1. Up Cx0x0 is the matrix of covariance of the coefficients
sampling is the process of lengthening a signal component by
inserting zeros between samples. Cx0nd = Cndx0T is the covariance matrix between coefficients to
estimate and the already measured coefficients
Training image – A model of scale dependency Cndnd is the covariance matrix between already
The idea of modelling the scale dependency of an attribute from measured coefficients
a training image was first introduced in geostatistics by A realisation of the conditional random vector is given by
Guardiano and Srivastava (1992). The training image describes (Rubinstein, 1981):
the geometrical facies patterns believed to be present in the
subsurface and reflect a prior geological/structural concept. −1
Zxa01nds
H 1 , D 1 ,V 1
≈ C xondsC ndsnds Z + Lx 0 nds e (2)
The needs for a training image in the proposed DWT
simulation comes from three sources:
where:
1. The scale dependency may not be present in the finite set of
measured points. e is a vector of four independent standard Gaussian
2. Discrete wavelet transform works only on full and regular values
grids.
Lx0|nd is the lower triangular matrix of the Cholesky
3. In many cases, the specialist has a clear understanding of decomposition of C xa01ndH 1 , D 1 ,V 1
the global features present in the attribute under interest.
For example, it can be a geological model inferred from If only a subset nds of the nd conditioning data is considered
indirect geophysical measurements or from geological based on the neighbourhood of x0, then applying screen effect
knowledge. approximation, Equation 2 becomes:
To summarise, DWT allows computing the directional
spectro-spatial characteristics of a training image. DWT and Zxa01nds
H 1 , D 1 ,V 1 −1
≈ C x 0 ndsC ndsnds Z + Lx 0 nds e (3)
IDWT is a bijonctive process without loss. Conditional simulation
of the wavelet coefficients will allow generating set of wavelet
coefficients resulting in an image that should have the same where Zxa01nds
H 1 , D 1 ,V 1
is a 4 × 1 vector of containing the simulated
wavelet domain characteristics as the original training image.
values of Zxa01 , ZxH01 , ZxD01 and ZVx 01 at x0. They all respect the auto-
In the next section, we describe how to cosimulate the wavelet
coefficients in the wavelet space and back transform to the and cross-covariances so as the conditioning data.
original space to produce conditional simulations of the attribute
of interest preserving its scale dependencies. Intra- and interscale covariance
CONDITIONAL SIMULATIONS WITH DISCRETE To enhance the practical application of the method, the full
WAVELET TRANSFORM IN TWO DIMENSIONS modelling of DWT coefficient auto- and cross-covariances is
replaced here with the hypothesis of intrinsic correlation. Then
In this section, a conditional simulation algorithm of wavelet the cross-covariance is equal to simple correlation (Chilés and
coefficients is presented. First, Sequential Gaussian CoSimulation Delfiner, 1999). Two cases need be considered, intra- and
(SGCS) (Verly, 1994) is briefly outlined and then, the algorithm is
interscales, as the scales have different supports. The intrascale
explained in detailed for wavelet coefficient simulation.
coefficients are collocated and are on the same support, hence,
Sequential Gaussian cosimulations of discrete the simple correlation r is straightforward. For example, the
correlation between ZH1 and ZV1 with zero mean is given by:
wavelet transform coefficients
In the proposed method, the approximation and the detail coeffi- cov( Z H 1 , ZV 1 )
cients are treated as regionalised covariables. The cosimulation r = (4)
Z H 1 , ZV 1
algorithm must reproduce the distribution and the variogram of var( Z H 1 )var( ZV 1 )

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 93
E GLOAGUEN and R DIMITRAKOPOULOS

where cov and var stand for covariance and variance, respectively. CASE STUDY
In the case of interscale, the variables belong to two different
scales having different spatial sampling. Similarly to the intrascale In order to explore our approach, the method presented above is
model, the cross-covariance between the inter-scale coefficients is applied to a synthetic case study where both geophysical and
computed using the simple correlation but spatial sampling has to geological data are known. Geological data (primary one)
be taken into account. The spatial relation between the wavelet corresponds to a channelled porosity field and the geophysical
coefficients belonging to different scales is represented by boxes data (secondary soft data) corresponds to the acoustic impedance
in the diagonal coefficient scales in Figure 1. Each box represents field (AI) measured on the geological model. Note that AI is
the same area but, for example, the resolution of the box in D1 is computed by multiplying the seismic velocity times the density
four times the resolution of the box in D2. In order to compute the of the probed material. Obviously, in the real data set, only the
correlation using relation 3, the values inside the highest resolution geophysical data are available. Figure 3 shows the synthetic
grid are averaged at the under sampled grid resolution. geological model we want to retrieve and Figure 2 shows the
corresponding geophysical model. In order to be as close as
Conditioning possible to real data acquisition, we added a Gaussian noise on
the geophysical data shown in Figure 2. Figure 4 shows the
As previously presented, DWT is a linear decomposition of an histogram of the noise free and the noisy AI (Figure 2). The
image. This allows the conditioning to existing soft or hard data standard deviation of the noise free AI is 275 m/s, whereas it is
to be implemented within the above simulation algorithm. 410 m/s for the noisy data. As can be seen on the scatter plot
In the soft conditioning case, the conditioning data (usually between noisy AI and the porosity field in Figure 5 the relation
geophysical data that shows systematic coverage over the whole between AI and porosity is weakly non-linear especially for
studied area, but with a poor resolution), measured at large scale in porosity below 25 per cent.
a regular grid, influence only the regional trend of the variable to
be simulated. In the case of wavelet coefficients simulation, soft
conditioning is obtain from converting all the approximation
coefficients of the secondary variable as conditioning data for the
approximation coefficients of the primary. To convert the
secondary data wavelet coefficients into primary ones, we produce
a geological analogue that represents the different coarse scale
textures present in the studied ground. This geological model is
translated in terms of geophysical properties using known
petro-physical relationships or in situ calibrated relations typically
obtained from borehole logs and cores. Both geological and
geophysical analogue are subjected to wavelet transformation.
Geophysical measurements give information on the large-scale
features of the ground and following this idea, focus is placed only
on the coarsest level of the wavelet coefficients. Statistical
relationships between the coarsest wavelet coefficients of both
geological and geophysical analogues are then studied. As noted
in the case study that follows, one of the most interesting
observations made during data analysis is that the weak
non-linearity between both analogues in the original space
becomes highly linear at the coarsest wavelet space. This occurs
principally because of the physics of the phenomenon – that is the FIG 2 - Measured acoustic impedance.
geophysical data are coarse scale representation of the geological
model. Due to the strong linearity of the relation between coarse
wavelet coefficients of both geophysical and geological analogues
and its ease to incorporate hard data, cokriging is used to infer
‘true’ geological coarse wavelet coefficients based on measured
geophysical wavelet coarse coefficients.

Quantification of the quality of the simulation

In this section, a quantitative way to quantify the quality of the
simulation is discussed.
The Julesz’ conjecture (Julesz, 1962) states that there exists a
set of statistical measurements such that two images are
perceptually indistinguishable if and only if they are drawn
from the same random function matching those statistics. In the
meantime, Julesz constructed examples, by creating pairs of
textures with the same nth-order pixel statistics that were
strikingly different in appearance. Also, it would be unlikely,
that a pair of images that produce identical statistical estimates
looks very different. In the present paper, the main
characteristics of the wavelet based simulations are that they
have to respect the covariances, the cross- covariances between
wavelet sub-bands and any conditioning data. Consequently, FIG 3 - Real porosity model.
performance of the method will be tested by applying DWT on
the simulated images and by computing their covariances and
cross-covariances of the DWT coefficients. Simulated images Based on our knowledge of the geology, we build a geological
wavelet coefficient statistics should reproduce the same analogue (Figure 6). The geological analogue differs from the
statistics as the training image wavelet coefficients. real field (Figure 3) by both the orientation of the channels

94 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL WAVELET BASED SIMULATION OF NON-STATIONARY GEOLOGIC FIELDS
FIG 4 - Noisy and noise free measured acoustic impedance.
FIG 5 - Scatter plot between acoustic impedance and porosity.
FIG 6 - Analogue porosity field.
(Figures 6 and 7) and histogram (Figure 8). A geophysical AI
field is computed from the analogue porosity field using known
petro-physical relationships (Pride, 1994). The corresponding
analogue AI field is shown in Figure 7.
Advances in Orebody Modelling and Strategic Mine Planning I
FIG 7 - Analogue acoustic impedance computed on porosity field.
As shown on Figure 1, third level wavelet decomposition is
applied on the measured geophysical data and on both geological
and geophysical analogues (Figure 8). The choice of the wavelet
decomposition scale is physically obtained when the correlation
between the geological and geophysical wavelet approximation
analogue is maximised. This approach allows the fusion of the
geological and geophysical data at the appropriate scale. Figure 9
shows the scatter plot between approximation coefficients of the
AI and porosity analogue. The relation between both data is now
very linear in the wavelet space and the correlation is as high as
0.981. As a comparison, the correlation between the measured AI
and its corresponding porosity is 0.656 whereas the correlation
between their wavelet approximation coefficients is 0.956. This
is due to the fact that geophysical data carries information on the
coarse scale of the probed ground. The approximation coefficients
of the porosity are cokriged using the modelled variogram of the
measured AI, the correlation coefficient between analogue AI and
porosity and the intrinsic hypothesis.
Finally, the variograms and cross-variagrams of each
sub-bands belonging to the same level or not are modelled. Using
the full linear model of corregionalisation and the porosity
approximation coefficients computed from the measured AI, the
porosity wavelet coefficients are cosimulated, starting from the
coarsest scale to the finest. When all the coefficients of all scales
are cosimulated, the wavelet coefficients are back-transformed in
the original space using IDWT. Figure 10 shows one simulated
porosity field. Compared to the real model shown in Figure 3, the
shape and intensity of the porosity of the background and within
channel is reasonably well reproduced except for some
numerical artifacts essentially due to the discretisation of the
analogues. Also, the correlation between the simulated porosity
field and the porosity model shown in Figure 3 is around 0.9. In
addition, even if the measured AI were really noisy, the
algorithm acts as a fine scale low-pass filter. It means that each
scale of an image can be filtered at its given scale. Finally, Figure
11 shows the e-type of 20 simulations. All the random numerical
artifacts due the discretisation vanished.
CONCLUSIONS
In this paper, it is demonstrated that DWT coefficients analysed
with well-known geostatistical methods can generate simulations
using modelling of the wavelet coefficient statistics of the field to
simulate. It can be shown that the wavelet scale statistics are
reproduced at least in terms of inter- and intrascale cross-
covariances. This proves that some scale dependencies are
reproduced in the simulated images. Realisations generated using
Spectrum Series Volume 17 95
E GLOAGUEN and R DIMITRAKOPOULOS

FIG 9 - Scatter plot between the analogue acoustic impedance

approximation coefficients of Figure 7 and the analogue porosity
approximation coefficients of Figure 6.

FIG 10 - One simulated porosity field.

FIG 8 - Three-level wavelet decomposition of: (a) measured

acoustic impedance (Figure 2), (b) porosity analogue (Figure 6)
and (c) acoustic impedance analogue (Figure 7). The
approximation coefficients are in the upper left box.

the proposed method reproduce the non-stationarity observed on

the geophysical data, the conditioning data, the wavelet
coefficients and their spatial dependencies. The coarse to scale
simulation approach allows simulating using both hard and soft
data and both are fitted at their appropriate scale. The methods still
work even if the analogue shows different spatial anisotropy in the
analogue. This property prevents the use of non-friendly rotation
matrices. The methodology can be easily extended to more than
one covariate geophysical data. The only change from the
proposed example is the cokriging step of the approximation
coefficient that becomes an n-variable cokriging instead of two FIG 11 - E-type of 20 realisations.

96 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CONDITIONAL WAVELET BASED SIMULATION OF NON-STATIONARY GEOLOGIC FIELDS
variables. A limitation of the method is that the training image has
to contain the same wavelet statistics as the field to be simulated.
Hence, the difficulty is to find appropriate training images. In
other hand, in the soft conditioing case, the method doesn’t require
the training image to have the same trend direction.
ACKNOWLEDGEMENTS
Thanks are in order to the industry members of the COSMO
Laboratory: AngloGold Ashanti, Barrick, BHP Billiton, De
Beers, Newmont, Vale and Vale Inco, as well as NSERC
(Discovery Grant 239019), the Canada Research Chairs Program
and CFI are gratefully acknowledged.
REFERENCES
Arpat, G and Caers, J, 2004. A multiple-scale, pattern-based approach to
sequential simulation, in Geostatistics Banff 2004, Springer,
Dodrecht (eds: O Leuangthong, and C V Deutsch), pp 255-264.
Chilès, J-P and Delfiner, P, 1999. Geostatistics, Modelling Spatial
Uncertainty, 385 p (Wiley: New York).
Choi, H and Baraniuk, R, 2001. Multiscale image segmentation using
wavelt domain hidden Markov models, IEEE Transactions on Image
Processing, vol 10, pp 1309-1321.
Crouse, M S, Nowak, R D and Baraniuk, R G, 1998. Wavelet-based
statistical signal processing using hidden Markov models, IEEE
Transactions on Signal Processing, 46(4):886-902.
Daly, C, 2004. Higher order models using entropy, Markov random fields
and sequential simulation, in Geostatistics Banff 2004 (eds: O
Leuangthong, and C V Deutsch), pp 215-224 (Springer: Dodrecht).
Fan, G and Xia, X-G, 2003. Wavelet-based texture analysis and synthesis
using hidden Markov models, IEEE Trans on Circuits and Systems –
Fundamentals, Theory and Applications, 50(1):106-120.
Flandrin, P, 1992. Wavelet analysis and synthesis of fractional Brownian
motion, IEEE Trans on Information Theory, 35:197-199.
Gloaguen, E and Dimitrakopoulos, R, 2009. Two-dimensional conditional
simulations based on the wavelet decomposition of training images,
Mathematical Geosciences, 41(6)679-701.
Goovaerts, P, 1997. Geostatistics for natural resources evaluation, 483 p
(Oxford University Press).
Advances in Orebody Modelling and Strategic Mine Planning I
Guardiano, F and Srivastava, R M, 1993. Multivariate geostatistics:
beyond bivariate moments, in Geostatistics Troia (ed: A Soares),
92:133-144.
Julesz, B, 1962. Visual pattern discrimination, IRE Trans Info Theory, IT,
volume 8, pp 84-92.
Kumar, P and Foufoula-Georgiou, E, 1997. Wavelet analysis for
geophysical applications, Reviews of Geophysics, 35(4):385-412.
Mallat, S, 1989. Multifrequency channel decompositions of images and
wavelet models, IEEE Transaction on Acoustic Speech and Signal
Processing, 37:2091-2110.
Mustapha, H and Dimitrakopoulos, R, 2010. Geologically enhanced
simulation of complex mineral deposits through high-order spatial
cumulants, in Advances in Orebody Modelling and Strategic Mine
Planning I (ed: R Dimitrakopoulos), pp 309-320 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Portilla, J and Simoncelli, E P, 2000. A parametric texture model based
on joint statistics of complex wavelet coefficients, Int J of Comput
Vision, 40(1).
Pride, S, 1994. Governing equations for the coupled electromagnetics and
acoustics porous media, Physical Review B, 50:15678-15696.
Rubinstein, R Y, 1981. Simulation and the Monte-Carlo method, 272 p
(Wiley: New York).
Strebelle, S, 2002. Conditional simulation of complex geological structures
using multiple-point statistics, Mathematical Geology, 34:1-21.
Tjelmeland, H and Eidsvik, J, 2004. Directional Metropolis: Hastings
updates for posteriors with non linear likelyhood in, Geostatistics
Banff 2004, pp 95-104 (eds: O Leuangthong and C V Deutsch)
(Springer: Dodrecht).
Tran, T, Mueller, U A and Bloom, L M, 2002. Multi-scale conditional
simulation of two-dimensional random processes using Haar
wavelets, Proceedings of the GAA Symposium, pp 56-78.
Verly, G W, 1994. Sequential Gaussian cosimulation: A simulation
method integrating several types of information in, Geostatistics
Troia, 1992 (ed: A Soares), pp 85-94 (Kluwer Academic Publishers:
Dordrecht).
Zhang, T, Switzer, P and Journel, A, 2006. Filter-based classification of
training image patterns for spatial simulation, Mathematical Geology,
volume 38, no 1, pp 63-80.
Spectrum Series Volume 17 97
 HOME

Optimal Life-of-Mine Scheduling for a Bauxite Mine

M Zuckerberg1, J van der Riet2, W Malajczuk3 and P Stone4

ABSTRACT combination of day/night and summer/winter. Also, over the

This paper describes a new life-of-mine planning optimisation software
mine life, a number of critical decisions need to be made
tool named Bodor (Boddington Optimiser) which has been developed for regarding the timing of investment in new crushing and overland
BHP Billiton’s Boddington Bauxite mine in South-Western Australia. conveyor infrastructure, or the merits of recycling existing
Bodor minimises pretax net present cost (capital and operational) over a crushers. Traditionally, mine planners have used some sensible
specified mine life, and is applied to a mine model consisting of bauxite heuristics to guide a scheduling model implemented in a popular
pods predesigned to a fixed cut-off grade, directly feeding a refinery with commercially available mine planning and scheduling tool.
bauxite that meets grade and throughput targets in each period over the These heuristics include delaying new crusher and conveyor
life-of-mine. Bodor has been very successful in its application to the infrastructure capital expenditure as long as possible and
Boddington mining operation, producing a new life-of-asset mine plan
which delivers a significant reduction in net present cost. This value add minimising operating haulage costs in obvious ways.
is achieved primarily through a better timing of orebody exploitation Unfortunately, these heuristics are not always reliable due to the
which optimally trades-off lower haulage costs against some capital costs natural trade-off between capital and operating costs, viz: by
that are brought forward. A core aspect of Bodor’s utility lies in delaying the expenditure of capital associated with the opening
managing various complex environmental constraints in determining the of new envelopes, the haulage costs will unavoidably increase as
optimal extraction schedule. These constraints lead to minimising noise the cheaply recoverable pods close to the existing crusher are
disruption to local landholders through accounting for seasonally progressively exhausted. Moreover, the confounding issues of
prevailing winds and day/night access controls, maximising the resource
meeting blend targets at the refinery gate and respecting
life through determination of the best blend strategies, and optimising the
mining truck fleet usage by balancing day and night deployment. environmental constraints add further complications. These
environmental constraints relate to the seasonality of prevailing
winds, the close proximity of landholders to some of the pods
INTRODUCTION (see Figure 3), and the desirability of opening, extracting, closing
The Boddington mine lies just south-east of Perth in and rehabilitating/remediating pods within a reasonable time
South-Western Australia and is a relatively low grade – average frame. Examples of the rehabilitation process are shown in
alumina grade of 31 per cent – but long-life bauxite resource, Figure 4.
which supplies the Worsley alumina refinery near Bunbury. It is a
joint-venture operation, 86 per cent owned by BHP Billiton,
ten per cent by Japan Alumina Associates and four per cent by
Sojitz Alumina. The bauxite ‘pods’ that constitute the resource of
some 700 million dry tonnes are spread over a lease area of
2700 square kilometres, and are clustered into several distinct
mining envelopes. Currently, only one of these envelopes is mined
and sent to the refinery via a 51 km overland conveyor. The Envelope 5
Boddington life-of-mine plan calls for the eventual development,
opening, extraction and remediation of pods in all mining
envelopes. Each envelope will be serviced by a crusher, with the
crushed ore being transported to the refinery via a network of
overland conveyors. Overburden and waste is stockpiled locally
and then used to remediate the shell of the exhausted pod. This
configuration, which is explained in more detail in the next section Envelope 4
(Bodor Asset Model), is illustrated in Figure 1 – the mining and
conveying operation is shown in Figure 2. Envelope 3

Mine planners at Boddington face a very complex task in

calculating the best extraction sequence for the resource over the Overland Conveyor
life-of-mine. Not only do refinery feed blend targets have to be
Crusher
met in each period, but the management of the mining truck fleet
is cost critical, given that many pods in the deposit can only be Envelope 2 Bauxite Pod
mined in the daytime, some in summer only and some in any

1. Principal Scientist, BHP Billiton Resource and Business

Optimisation, 180 Lonsdale Street, Melbourne Vic 3000.
Email: mark.zuckerberg@bhpbilliton.com Refinery

2. (Deceased) Formerly with: BHP Billiton Bauxite Alumina Technology

Centre, Gastaldo Road, Worsley via Collie WA 6225.
3. MAusIMM, Superintendent Mine Planning, Worsley Alumina Pty Envelope 1
Ltd, Boddington Bauxite Mine, Gastaldo Road, Worsley via Collie
WA 6225. Email: wally.malajczuk@bhpbilliton.com
4. Manager Optimisation R&D, BHP Billiton Resource and Business
Optimisation, 180 Lonsdale Street, Melbourne Vic 3000.
Email: peter.m.stone@bhpbilliton.com FIG 1 - Bauxite mine asset model topology.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 101
M ZUCKERBERG et al

Bodor (Boddington Optimiser) has been developed in-house at

BHP Billiton’s Melbourne Technology Centre to enable
Boddington’s mine planners to calculate a resource extraction
sequence over the life-of-mine that minimises net present cost
(capital and operational), meet all bauxite blend targets at the
refinery front gate, respect environmental and operational
constraints, and allow smooth utilisation of trucking resources. It
is a software product with a graphical user interface and a
mixed-integer-linear-programming optimisation engine, and is
used directly by Boddington mine planners to evaluate
life-of-mine plans rapidly – a single optimal life-of-mine plan
can be calculated overnight. In this paper, we will outline the
capability of Bodor and show how it has been used by
Boddington mine planners to develop optimised life-of-mine
plans. In the last section, we discuss several enhancements to
Bodor already in the development pipeline.

BODOR ASSET MODEL

FIG 2 - Photos of the Boddington mining operation, showing
truck/shovel excavation of shallow surface pods and the overland
conveyor to the Worsley refinery.
The centrepiece of the Bodor asset model is the ‘pod’, which is a
distinct body of ore lying close to the surface and of modest size
– usually less than two million tonnes. Bodor is specifically
designed for situations in which there are numerous pods that are
relatively small in size – of the order of several hundred – so that
no single pod would require more than approximately three years
to mine. Bodor consequently focuses on the sequence in which
pods are to be excavated, rather than on the block extraction
sequences of the individual pods. At Boddington, some of the
pods are composed of a collection of ‘subpods’, effectively
simple mining phases. Subpods must be developed in a
designated sequence as illustrated in Figure 5. Each subpod is
assumed to be homogeneous in all respects and hence can be
extracted in the optimised schedule in any proportion. Subpods
have a physical mining precedence, one to another. The
precedence rules are that subpods are numbered in their strict
order of possible extraction; viz, subpod 2 can only be extracted
after subpod 1 and so on. Figure 1 illustrates that each pod is a
member of a mining envelope. All pods in a given envelope
deliver their ore to a particular associated crusher. There is a
one-off capital cost associated with ‘opening’ an envelope. There
are also one-off capital costs associated with building the crusher
(or relocating it from another envelope) and the necessary
conveying system used at the envelope. No mining can occur
within an envelope before the envelope is opened. The cost to
open an envelope will depend on which and how many envelopes
have been previously opened and closed, as different envelopes
will share some infrastructure.
With regard to conveyers, at Boddington all ore mined from an
envelope undergoes primary crushing at the envelope’s local
crusher and is then conveyed to a secondary crushing point and
eventually on to the refinery. The conveying system need not
connect each envelope directly to the refinery. Rather, it is a
directed tree network of arcs (conveyer increments) with a root
node at the refinery and nodes representing the envelopes, such
that the ore of each envelope is conveyed via the arcs (conveyer
increments) of the tree on its way to the refinery. Specifically,
each envelope has a single downstream path leading to the
refinery and thus this entire path of arcs (conveyer increments)
must be constructed before the envelope may open.

BODOR OPTIMISATION PROBLEM

FIG 3 - Position of noise-sensitive residences in relation to current
mining operations at Boddington.
Objective
The Bodor objective is to calculate a feasible subpod extraction
schedule and set of exploitation decisions which will minimise
the net present cost and deliver the target tonnage and blend
quality to the refinery gate in each period whilst absolutely
obeying all constraints over a predetermined mine life. It was

102 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I

FIG 4 - Examples of land and forest rehabilitation at Boddington.
chosen to minimise net present cost, rather than maximise net
present value, because a representative and fair revenue model
for the bauxite at the refinery gate was not available.
Exploitation decisions
Bodor is designed to answer such strategic life-of-mine planning
questions as:
• the determination of mining envelope’s opening and closing
times,
• whether crushers should be constructed or redeployed from
one mining envelope to another mining envelope,
• whether or not the conveyer infrastructure for one of the
mining envelopes may be redeployed at a different mining
envelope, and
• the determination of optimal truck fleet sizes and optimally
smooth allocations of trucks to the various classes of pods in
each mining envelope.
Inputs
The pod model will contain all grade and cost information,
relating to the bauxite in-the-ground. In terms of costs, the pod
model should separately list the following:
Sub-Pod 4 Sub-Pod 3
FIG 5 - Bodor subpod topology.
Advances in Orebody Modelling and Strategic Mine Planning I
OPTIMAL LIFE-OF-MINE SCHEDULING FOR A BAUXITE MINE
• costs that are fixed and borne at the start of pod excavation,
• costs that are fixed and borne in the development period
before the start of pod excavation,
• costs that are fixed and borne at the conclusion of pod
excavation, and
• costs that are charged in proportion to the amount of the pod
extracted (able to be represented as $/tonne extracted).
The planner must also provide Bodor with the asset model
topology, as well as all relevant constraints and targets applying
over the prescribed mine life.
Constraints
The following constraints can be optionally applied by the
planner in the determination of Bodor’s optimised schedule:
• Tonnage: target tonnage of alumina or bauxite delivered to
the refinery gate per period.
• Blend: upper and lower limits to the percentage of reactive
silica, total silica, extractable sulfate, extractable oxalate and
carbon in bauxite delivered to the refinery gate per period.
• Maximum pod open time.
• Subpods: for larger pods, subpods are mining phases that are
separately extractable and must be taken in a strict order.
Sub-Pod 2 Sub-Pod 1
Spectrum Series Volume 17 103
M ZUCKERBERG et al
• Grade control drilled pods: those pods that have already been
grade controlled can be given precedence in the schedule.
• Earliest start and latest finish.
• Maximum mining rate.
• Maximum conveyor capacities.
• Noise and dust: these are day/night and summer/winter
restrictions, which are applied to manage both noise and dust
pollution and to limit the general impact of mining
operations on adjacent landholders. Every hour available for
mining is classified either as ‘day’ or as ‘night’, and either as
‘summer’ or ‘winter’. Time can thus be divided into four
disjoint categories of summer day, summer night, winter day
and winter night. Pods may be designated as available for
mining during only some of these four times. For example, a
pod may be ‘day only’ ‘summer only’, or ‘summer day and
winter night only’, etc. These constraints serve essentially to
limit the amount of ore that may be extracted from such pods
in each period.
• Lead time: the user will specify the number of months
required for the construction or relocation of conveyers and
crushers. No initially closed envelope will be allowed to open
before the requisite lead time elapses for constructing its
infrastructure. Furthermore, no envelope will be allowed to
open with a relocated crusher or conveyer before the requisite
lead time elapses after the closure of the envelope from which
the crusher or conveyer is being relocated.
• Smooth trucking: smooth and consistent utilisation of truck
resources can be enforced.
Outputs
Bodor’s optimiser engine will output an assignment of a fraction
between zero and one to each subpod and period, indicating the
fraction of the ore in the corresponding pod that is to be excavated
in the corresponding period. For example, Bodor may output the
number 0.4 to subpod I period J, which will indicate that 40 per
cent of subpod I is to be excavated during period J, but Bodor will
not identify which 40 per cent of the ore ought to be extracted.
Bodor assumes the material in the subpod to be homogenous, so
that the 40 per cent of the subpod that was extracted will be
assumed to possess the same grade parameters as the subpod as a
whole. Periods can be of any length, practically from one month to
several years as required by the planner.
The optimiser will output a comprehensive schedule of all pod
openings and extractions, as well as of the openings and closings
of each of the envelopes and the construction dates of each of the
conveyer increments. For each envelope it will indicate whether
the envelope is to be configured with a new or a relocated
crusher. Bodor will also determine the best truck fleet size for
each lease period, and output a comprehensive truck assignment
schedule that indicates how many trucks were assigned to each
of the envelopes during each truck assignment interval.
BODOR CONFIGURATION
Bodor is designed around the CPLEX optimisation suite by
ILOG Inc (ILOG (2007)), and is deployed for use by operations
mine planners on high-end laptop computers. It consists of three
major components:
1. a user interface, where planners configure the mine asset
model and assign parameter values to the model and for the
optimisation procedure,
2. the asset model, realised as a mixed-integer-linear-
programming (MILP) problem using CPLEX Concert
technology, and
3. the standard CPLEX MILP optimisation engine.
104 Spectrum Series Volume 17
Although all parameters are configured via the graphical user
interface, Bodor also provides a batch-processing facility –
planners may configure a number of optimisation scenarios and
submit them to the batching utility for sequential processing
overnight.
Two-phased approach
Mine planners at Boddington currently calculate life-of-mine
schedules on the basis of 18 single month periods, followed by
14 quarterly periods, followed by between 25 and 35 yearly
periods. This is a very large number of periods for a mine
planning optimisation problem, and with no guidance as to
when envelopes and sectors will open and close, this problem in
Bodor is not currently tractable. A 36 year schedule whose
periods are 18 months each, however, is indeed tractable. This
suggests that a two-phased approach can be used to render the
whole optimisation problem tractable. In the first phase, larger
period lengths with some constraint relaxations are used to
determine approximate opening and closing times for each pod
and in the second phase the full optimisation problem is solved
for the shorter period lengths by applying relaxed earliest start
and latest finish constraints as informed by the first phase
results. Thus, the primary purpose of the first phase of
optimisation is to bind the opening and closing times of the
envelopes and pods to within narrow, but not over-constraining,
limits. We have tested Bodor extensively to confirm that the
loss in value due to the implementation of this two-phase
approach is minimal, in the order of fractions of one per cent of
the total net present cost.
BODOR PERFORMANCE
Bodor has been applied to the 30-year plan at the Boddington
Bauxite mine for the existing 3.5 Mt/a (alumina) operational
configuration. The Bodor schedule was compared to the existing
benchmark schedule, which was developed using manual guidance
of the XPAC scheduling software. In the development of the
benchmark, the principle used to guide the XPAC scheduler was to
delay major capital expenditure as long as possible. In the end, the
benchmark schedule did not respect all constraints in all periods
absolutely. However, it represented the best schedule that mine
planners could construct with existing commercially available
tools. In contrast, the Bodor schedule does respect all constraints
in all periods and furthermore delivered a net present cost that is
five per cent lower than that of the benchmark. The majority of
this value is delivered by implementing a better mining strategy.
Figure 6 shows that Bodor brings forward the opening of new
envelopes and the associated capital charge, but balances this
through smaller truck fleets, reduced truck cycle times and more
efficient utilisation of the truck fleet (Figure 7).
CONCLUSIONS
The success of Bodor, applied at the Boddington Bauxite Mine,
demonstrates the value of state-of-the-art optimisation tools in
the life-of-mine planning domain. Using a powerful MILP solver
such as CPLEX combined with appropriately designed mining
models and careful implementation of a two-phase optimisation
strategy has allowed us to develop life-of-mine plans that respect
complex shorter-term operational constraints on a monthly basis
and also deliver a minimum net present cost bottom line.
REFERENCES
ILOG Inc, 2007. ILOG CPLEX v10. Available from:
<http://ilog.com/products/cplex/>.
Advances in Orebody Modelling and Strategic Mine Planning I
 OPTIMAL LIFE-OF-MINE SCHEDULING FOR A BAUXITE MINE

Bodor
Exploitation Seq. by Envelope

ENVELOPE 1 TONNES ENVELOPE 2 TONNES ENVELOPE 3 TONNES ENVELOPE 4 TONNES ENVELOPE 5 TONNES ENVELOPE 6 TONNES ENVELOPE 7 TONNES

100%

80%

% Mining Envelope
60%

40%

20%

0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Period

Benchmark
Exploitation Seq. by Envelope

ENVELOPE 1 TONNES ENVELOPE 2 TONNES ENVELOPE 3 TONNES ENVELOPE 4 TONNES ENVELOPE 5 TONNES ENVELOPE 6 TONNES ENVELOPE 7 TONNES

100%

80%
% Mining Envelope

60%

40%

20%

0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

Period

FIG 6 - Boddington envelope exploitation schedules.

Bodor
Noise Sensitive Tonnage Trend

SUMMER DAY TRUCK HOURS SUMMER NIGHT TRUCK HOURS WINTER DAY TRUCK HOURS WINTER NIGHT TRUCK HOURS

100%
% Total Utilised Truck Hours

80%

60%

40%

20%

0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Period

Benchmark
Noise Sensitive Tonnage Trend
SUMMER DAY TRUCK HOURS SUMMER NIGHT TRUCK HOURS WINTER DAY TRUCK HOURS WINTER NIGHT TRUCK HOURS

100%

90%
% Total Utilised Truck Hours

80%

70%

60%
50%
40%

30%

20%
10%

0%
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30
Period

FIG 7 - Boddington truck utilisation.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 105
 HOME

Coal Mining in the Hunter Valley – A Strategic Mine Planning

Case Study
C Wharton1

ABSTRACT basal seams and mining extents on tonnage and NPV was also
undertaken. The model was then later used to plan plant
Coal presents many engineering challenges and opportunities for
optimisation in a changing economic environment. The imminent impact expansions to cope with new coal fields, as well as decisions on
of carbon taxation and cleaner coal technology will add to these location, including truck haulage and/or conveyor options.
challenges. Ash content and other physical attributes dictate what types of
coal are available; however, the ability to wash and change the coal COAL MODELLING
characteristics, together with bypass and blending, leads to interesting
tonnage versus value opportunities – sophisticated tools are required to Inputs to the model
model and optimise these calculations. Long mine lives present a
challenge to net present value based optimisation – this paper will discuss The inputs to the model are as follows.
this issue, along with alternatives. A recent review of the Hunter Valley
Operations managed by Rio Tinto Coal Australia Pty Limited (RTCA) is Phase files
used as a case study. RTCA has investments in several mines in the same
area which currently share rail and port facilities with other producers. Each independently scheduled area will be referred to as a phase –
This case study will discuss the many issues involved in these mines and Bengalla has four phases, HVO has 17 and MTW has six. As
show how a coordinated approach can improve the profitability and shown in Figure 1, coal phases are mined vertically, one strip at a
performance of the mines. time. The mining consists of removing waste overburden and then
removing the exposed coal seam. When the bottom of one strip is
INTRODUCTION reached, the operation advances to the next strip. The actual
implementation of this is more complex, allowing for slope
Rio Tinto Coal Australia Pty Limited (RTCA) manages the Coal layback and other equipment issues. The existing computer
& Allied (CAN) operations in the Hunter Valley, consisting of generated equipment schedules, consisting of pit, strip, bench,
Bengalla (BEN), Hunter Valley Operations (HVO) and Mount block and start time, were used to create, for each phase, a
Thorley Warkworth (MTW). The mines produce a range of coal time-based sequence of waste bcm, coal tonnage, ash and other
products for both the export and domestic markets, including qualitative data for use with the computer model.
thermal coal, semi-soft coking coal and low, medium and high
ash export coals. All export coal from these mines and other
producers in the region is transported by rail to the Port of
Newcastle, where it is loaded by Port Waratah Coal Services 5
overburden
onto ships. The rail and port facilities are currently capacity coal
interburden 1 6
constrained. coal
RTCA initiated a review of the way they do business that 2
would allow them to consider all aspects of their operations. 7
They proposed a three pronged attack. The first stage was to Benches
build a computer model that would produce a life-of-mine 3
schedule, then validate it against the MTW operations to ensure
that they could replicate the existing net present value (NPV).
MTW was chosen as it had the most complex operation, with 4
two coal plants and the ability to do one- and two-stage washing ips
and bypass. The second phase was to review all the value drivers Blocks Str
for the operation. The methodology and options explored may be
of interest, and show what can be achieved or uncovered when FIG 1 - Coal seam schematic.
the parameters are opened up for renewed appraisal. This paper
outlines some of the key analyses undertaken by the author in
this work. Other mines may benefit from the same type of Revenue
approach. The policies investigated included:
Coal price varies based on thermal energy and ash content. The
(a) selection of the coal ash/yield set point to maximise ash content also controls product groups:
revenue; • semisoft,
(b) determining the best wash plant to use for each coal • low,
seam, taking yield, processing and haulage costs into
consideration; and • medium and high ash export, and
(c) the application of equipment capital costs when changing • domestic.
optimisation sequence. Modelling these attributes allows the revenue to be calculated
Policies (a) and (c) are discussed in detail below – (b) is not for each coal class.
discussed further. A review of the effect of changes to pit limits,
Costs
1. Principal Consultant, Strategy Optimisation Systems Pty Ltd, Fixed and variable costs were allowed for mining equipment,
66 Rathmullen Quad, Doncaster Vic 3108. mining services, wash plant processing, general overheads, rail
Email: Chris.Wharton@stops.com.au haulage and port loading.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 107
C WHARTON

Constraints data. MTW is currently using a set point of 1.6. Table 1 shows
the total coal tonnage available when running the optimisation at
Equipment capacities, mining and wash plant throughputs, rail six other set point settings. As can be seen, the yield drops
and port limits were entered based on existing conditions. dramatically when the set point is lower than 1.5 – price
Practical development constraints related to pit access and lead differentials would have to be significant to make this
times were also included. The optimisation algorithm is worthwhile. Yield increases above 1.6 and it will be the ash and
discussed in the Appendix. qualitative values that determine whether the additional product
can be sold.
Outputs from the model
The key outputs are listed below. TABLE 1
Coal tonnage by wash plant set point.
Period schedule
Set point Domestic Export Total Tonnes diff
Shows policy setting(s), cash flow, quantities of waste by (Mt) (Mt) (Mt) (%)
equipment, coal and product tonnage, ash percentage and other 1.40 7.3 197 204 -26.3
quality attributes, by class, phase and equipment operating hours.
1.45 8.7 229 238 -14.0
Phase summary 1.50 10.1 262 272 -1.9
1.55 12.4 264 277 -0.2
More detailed break-down of scheduled quantities and attributes
by phase by period. 1.60 14.2 263 277 0.0
1.65 15.9 269 285 2.9
Constraints 1.70 18.0 280 298 7.6
Shows what is limiting the schedule for each period.

Investigation of options
Calibration runs
Existing mining plans were used to get the annual wash plant coal
tonnage inputs as a target for the calibration fixed sequence
schedule. The computer program was restricted to only find those
tonnages. Comparison of the calibration run with existing
schedules showed physicals within one per cent and NPV within
five per cent.
Initial optimisation runs
An optimisation was then run using only the plant capacity
constraints to see what other value might be generated. This may
appear unrealistic, but the computer program was run to see which
pits generated the best NPV and what it would then do with the
problem. If you don’t relax some constraints, then you run the risk
of getting a similar solution to your fixed schedule, as there is no
freedom or room to move. The optimisation showed a large
improvement in NPV; however, some of the resulting equipment
requirements exceeded practical pit capacity, given access and
dump space limitations. What it did highlight was the profitability
of the Battleaxe region by bringing it forward in the schedule. A
further series of optimisation runs were carried out, adding
dragline and shovel equipment limits to each of the pits, overall
equipment limits, and revised pit and phase sequence constraints,
until a satisfactory schedule was produced. This then became the
new benchmark for other option investigations. A similar situation
was found while calibrating the HVO mines where the Cheshunt
pit was shown to be a major driver of value. This led to a revision
of existing plans and the installation of more shovel capacity in the
Cheshunt pit.
Ash/yield trade off
Wash plants use a density set point to determine the amount of
ash in the final product. The coal is floated off – the heavier
material sinks and is rejected. You can obtain better quality
(lower ash) with a lower yield and this may be attractive if the
price increment is sufficient to offset the loss of tonnage. This is
particularly important when the capacity of rail and port
infrastructure is potentially constraining. To model this sort of
complexity you need ash yield data at a variety of set points.
Currently this level of detail is only available for the MTW seam
The ‘best’ set point policy will depend on the coal prices for
each of the coal categories, and plant and product constraints. If
there is little premium for a semi-soft product, then this material
may have more value being blended with low and medium ash
grades to achieve better overall value. Figure 2 shows the
life-of-mine total tonnage and average ash content, for each coal
class, for a fixed and a variable set point run. With the variable
set point run there is less total semi-soft material (-8.8 Mt)
and more low ash (9.1 Mt), high ash (12.1 Mt) and domestic
(3.3 Mt). The ash percentage, by coal class, tends to be slightly
higher with the variable set point, however, there is an increase in
coal yield of 15.7 Mt and an improvement in the NPV.
Pit limits and extents
A series of new pit models was produced based on five
margin-ranked shells for each of four different coal seam floors
and the existing case. Dragline horizons and configurations were
adjusted, depending on the seam floor depth, and all cases were
scheduled in a consistent manner and optimised. All the cases,
including the current case, were plotted in increasing NPV order
together with the total tonnage, as shown in Figure 3. The current
case is shown as data point A. In general, the shallower options,
while they had a smaller equipment footprint, failed to deliver a
reasonable NPV. The deeper options had more NPV, but also
delivered over twice the current tonnage and excessive mine lives
based on current extraction rates. The work was not in vain as
indicated by two potentially interesting cases (B and C).
Case B has an NPV that is almost the same as the current case,
but has a mine tonnage almost 20 per cent less than the current
case. Case C has a tonnage that is almost the same, but a higher
NPV. There might be additional boxcut costs required to integrate
it into the existing design, but the NPV increase is attractive
enough to warrant further investigation. Both these cases should
be scheduled in a detailed manner to confirm their NPV values
and schedules. There is also the opportunity to explore other
combinations of pit floors and extents for individual pits based on
the work done. There may be additional value from going to
different floors in the different pits.
Capital application
The computer optimisation should consider the capital required
if additional trucks or shovels are added to achieve revised
scheduling plans. This leads to more realistic schedules, and

108 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 COAL MINING IN THE HUNTER VALLEY – A STRATEGIC MINE PLANNING CASE STUDY

100 30%
90 27%

Mtonnes Product Coal

80 24%
70 21%

Product Ash
60 18%
50 15%
40 12%
30 9%
20 6%
10 3%
0 Exp Dom
0%
SS 1 SS 2 LAE 3 LAE 4 LAE 5 HAE 6 HAE 7 Dom 8
Byp BDT

Tonnes, Fixed SP 47.4 38.1 76.6 50.9 27.5 8.8 7.7 9.2 7.7 7.0
Tonnes, Variable SP 25.0 51.7 73.1 47.0 44.0 17.6 11.0 9.2 11.0 7.0
Ash, Fixed SP (rhs) 6.6% 8.6% 10.4% 12.5% 14.3% 16.8% 23.5% 11.1% 23.5% 15.0%

Ash, Variable SP (rhs) 6.5% 8.7% 10.5% 12.5% 14.4% 16.6% 24.3% 11.1% 24.3% 15.0%

Tonnes, Fixed SP Tonnes, Variable SP Ash, Fixed SP (rhs) Ash, Variable SP (rhs)

Product codes
SS semi-soft, LAE low ash export, HAE high ash export, DOM domestic

FIG 2 - Ash/yield – fixed versus variable set point.

160% 12000.0

140% NPV
10000.0
Tonnes
Current C
120%

B A 8000.0

Tonnes (m)
100%
NPV (%)

80% 6000.0

60%
4000.0
40%
2000.0
20%

0% 0.0
W 2 L 4. 2 65 6
C 1 1 18 T 26 3

.3 9 0
3

L .6 7 1
N 5 3 7 8
FJ t 8 4 2

M 16 3 T 37

B 2 L T 42
3 16 T 35

F 3 8 T 5

W U R 3 L 8 .2 T 6 9
B 2 L .1 5 7 6
C W 8 8 T 5

D G M D 1 1 7 .7 1 8 6

4 4 1 3 7

7
4 D 3 .7 1 6

W B F 2 L 3 .2 2 7

W W K 4 35 . T 5 3
C K 4 15. T 5 14

B Q 418 33. 9 17

11 0
Q 19 L 9 T 958
W Q 4 L 7. 60 2
U 5 0 3 7
Q 7 18 T 37
M M c 14 T 2

U 9 22 T 9

C K 17 37 T 44

B 1 45 7 T 989
C CU 010 L 2 3 T 91

55 T 00
W 1 L 3 1 T 23

0 45 T 65
rB 01 L 2 .1 T 860
B flee L 1 9 T 12

B 4 L 2 .6 8 3

T 16
C 4 1 L 2 .4 8 2 5
W 0 2 T 84
W U R 2 L .9 69
G 8 L T 3

2 2 L T 53

29
4
B re L 2 .8 T 9 5

N 1 3 T 4
C 0 L 1 .6 T 4 3 3
2 M 1 L8 T 2

6 2 T 0
G G t + L .9 1 6 6

C J 2 t L .8 3 8 4
4

B NQ 2 L .5 45
u r K 1 .6 1 2

G 8 L 1 . 5 43

ur 61 8 9 7
B 6 L .2 3

R L .9 0

FJ L 22 3

. 4
4

02 L 6 6
D D a p .2 3

4
C 6 L 5 .4 T

T
5
D 6 3

3
Q L .
F n 1

K 5 0
G J8 1 L

K 5 9
fle 4 L
B D8

W J

W 6

N 2
M

R
ur
F

ur R
D

J
M
G

N
u
M
G

FIG 3 - Alternative pit configurations.

stops equipment from being added and not adequately used.
While the NPV improvements may not be large, they are easy to
implement and have little impact on the running time. The NPV
will also be more realistic and agree more closely with
traditional assessments.
Hunter Valley integrated models
Once each mine site is calibrated it can be added into an
integrated model. This model maintains individual site mine,
plant and equipment limits, and adds potential port and rail
constraints. This means that alternative expansion plans can be
measured and assessed on a common basis. The model has also
been used to investigate plant expansions to cope with new coal
fields, and decisions on plant location, including truck haulage
and/or conveyor options.
Long life projects
The HVO mine has a long projected life which poses issues for
evaluations based solely on NPV. This problem is not unique to
coal, but applies to all mines with long lives. If you have several
alternatives with similar NPVs, what other criteria can you apply
to determine which is ‘better’? The reason for this problem is

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 109
C WHARTON

that the discount decay means that there is little contribution to Carbon tax issues
NPV once the mine life gets over 18 to 20 years. Figure 4 may
help to illustrate the problem. It shows the yearly discount factor Greenhouse gas emissions are created directly from exposed coal
for a ten per cent discount rate. The contributions at years ten, 20 seams and indirectly by the use of diesel and electricity during
and 30 are approximately 40 per cent, 15 per cent and five per the mining process. The coal products also create greenhouse gas
cent, respectively. If we assumed that the net revenues were emissions when used by customers for power generation and
evenly distributed over the mine life, then the first five years converting iron ore into steel. Garnaut (2008) has stated:
contributes approximately 40 - 45 per cent of the NPV, and by If the coal industry is to have a long-term future
year ten you have accounted for 60 - 70 per cent of the value. in a low-emissions economy, then it will have to
be transformed to near-zero emissions, from
10% Discount Rate
source to end use, by mid-century.
1 100%
0.9
RTCA has a published climate change action plan and is
currently supporting research into sequestration through their

Per cent contribution

0.8 80%
support of the CO2CRC Otway carbon capture project, and into
Discount factor

0.7
0.6
0.5
0.4
0.3
0.2
0.1
0 0%
0 10 20 30
Discount 10yr 20yr 30yr
FIG 4 - Discount rate and value contribution by period.
There are other qualitative measures that should also be
assessed over the mine life:
• plant throughput and head grades,
• peak mining rates and equipment use,
• capital requirements, and
• return on investment and payback periods.
You could also look at other metrics, such as capital intensity, or
product produced for different periods of the project, then choose
the case that had better values at the start of the project. A similar
approach is to break the mine down into Business Risk Periods
(BRP) (Steffen, 1997). The BRP is a period, usually seven or eight
years, over which the mine makes investment decisions and
expects to make a profit. It distinguishes risks associated with
business from those associated with geology. You break the mine
life into BRPs and plan within those BRPs. Project risks change
over time – this can be highlighted in different BRPs as shown in
Table 2, which is the author’s illustrative interpretation of the way
in which risk may change through the life-of-mine and the BRPs.
For example, price is not normally an issue at the start of the
project, as you may have contracts, hedging, or recent history to
guide you. Costs are reasonably well known and you should have
measured reserves to start with, so the geology should be low risk.
The more likely risks for start-up mines could be process
technology, initial plant performance or ramp up. For long life
mines, there might be the consideration of new technology or
product substitution.
60% technologies which reduce carbon dioxide emissions through
their support of the Coal21 fund.
40%
The computer models are capable of considering a range of
20%
future carbon priced scenarios. It is straightforward to add
another cost or range of costs into the optimisation cashflow
logic and see the financial impact of a carbon tax on the project
40 50 value, but there are several other questions that need to be asked:
40yr 50yr What is the appropriate impost? There have been several
suggested rates but no firm government policy as yet. How will it
escalate with time? You can run a series of cases to understand
the possible impact. Will any of the costs be passed on to the
customer? If they cannot, then it is important to determine where
the company is positioned in the industry cost curve, and what
the likely impact is on their viability, compared with their
competitors. Are customers likely to look for a cleaner substitute,
and does this have an impact on proposed mine life? Finally, are
there technologies available or being developed that may assist in
mitigating the carbon footprint, and at what cost and timeframe?
CONCLUSIONS
Reviewing and re-evaluating the value drivers for the project
turned out to be a worthwhile exercise. It did require a lot of
work setting up the data to support the analyses, but RTCA
gained insights into what was producing value and where they
could focus their current and future efforts. Each of the mines
benefitted from a review of the optimisation possibilities. The
runs revealed alternative sequencing arrangements or situations
where additional equipment added value. With a computer model
in place, RTCA gained the ability to analyse a lot of alternatives
in a speedy and consistent manner. This has helped them make
informed decisions regarding future expansion plans.
ACKNOWLEDGEMENTS
The author would like to thank RTCA for allowing the
publication of this paper. M Scott, the general manager of mine
planning, has been particularly supportive through all the
computer modelling.

TABLE 2
Business risk periods.

Component BRP 1 BRP 2 BRP n

Equipment/technology • Is the equipment proven? • New competitors • Can you take advantage of new
• Can you start small and ramp up? • Position on cost curve technology?
Price • Not normally an issue as the grades • Are prices rising or falling? • Critical for incremental costs in the
are better in the inner pits outer pit shells
Cut-off • Can you use variable cut-offs? • Are there any stripping hurdles? • Can you keep the mill full?
• Can you access the orebody?
Payback • Payback on initial investment • Payback on next investment • Payback on next investment

110 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 COAL MINING IN THE HUNTER VALLEY – A STRATEGIC MINE PLANNING CASE STUDY
REFERENCES
Australian Coal Association, 2004. Coal 21 National Action Plan
[online]. Available from: <http://www.coal21.com.au/Media/COAL
21_Action_Plan.pdf>.
Cooperative Research Centre for Greenhouse Gas Technologies
(Co2CRC), 2008. 2008 Annual Report [online]. Available from:
<http://www.co2crc.com.au/dls/annreps/08/ResearchPrograms_CO2
CRC_Annual0708.pdf>.
Garnaut, R, 2008. Garnaut Climate Change Review Final Report – 30
September 2008 (full version) [online]. Available from: <http://www.
garnautreview.org.au/> [Accessed: 27 October 2008].
Lane, K F, 1988. The Economic Definition of Ore: Cut-off Grades in
Theory and Practice (Mining Journal Books Limited).
Rio Tinto Coal Australia Pty Limited, 2006. RTCA Climate Change
Action Plan October 2006 [online] Available from: <http://www.
riotintocoalaustralia.com.au/documents/2_-_Climate_Change_Action
_Plan.pdf>.
Rio Tinto Coal Australia Pty Limited 2007. 2007 RTCA climate change
review [online]. Available from: <http://www.riotintocoalaustralia.
com.au/documents/2007Review_CCAP.pdf>.
Steffen, O, 1997. Planning of open pit mines on a risk basis, The Journal
of the South African Institute of Mining and Metallurgy, March/April
97:47-56.
Wooller, R, 2007. Optimising multiple operating policies for exploiting
complex resources – An overview of the COMET scheduler, in
Orebody Modelling and Strategic Mine Planning, second edition (ed:
R Dimitrakopoulos), pp 289-296 (The Australasian Institute of Mining
and Metallurgy: Melbourne).
APPENDIX
Optimisation algorithm
The optimisation algorithm uses dynamic programming with
iterative successive approximations to search for the maximum
NPV over the life of the mine, using defined project policies. The
policies currently include, but are not limited to, the sequencing
of the phases, the selection of the wash plant set point and the
application of equipment capital costs. There is no guarantee that
the solution will converge on the optimum, however, this
technique has been used successfully to improve project values.
Advances in Orebody Modelling and Strategic Mine Planning I
The formulation of the problem, using the Lane nomenclature,
is given by the equation:
Max ⎧ V '( R − r, T + t ) ⎫
V '( R, T ) = 0≤ r ≤ R ⎨c( t, r, w ) + ⎬
all ω ⎩ (1 + δ )t ⎭
where:
the maximum NPV (V) of a resource (R) at a time (T) is
calculated by finding the best strategy (ω) that maximises the
sum of the cash flow (c) for an increment of the resource (r) and
the maximum NPV of the remaining reserve.
The remaining resource cash flows are discounted (by δ) and
start after the increment of resource is mined (at time t). In this
formulation the maximum value (V) of the entire reserve and
remaining reserve is replaced by an estimate of the maximum
value (V’). The phase files define the mining sequence
increments. The policies will control what actions to take in each
period, the revenue and costs will determine cash flow and the
constraints will control how much of the resource is mined.
The resource is mined from the beginning, and an estimate of
the remaining value is determined by scheduling the resource,
using default properties, in an initial pass. These defaults may be
a breakeven cut-off grade, or in the case of coal, an assumption
that all uncovered coal will be used with a yield based on the
current wash plant settings. The algorithm then repeats the
calculations, optimising the operating policies using the previous
value estimate, until the requested number of iterations is
reached or no improvement in NPV is found compared with the
previous iteration. Typically, this process requires less than ten
iterations to converge on a solution. Wooller (2007) provides a
good description of all the inputs and operation of the computer
program. This algorithm uses a number of sequences in ‘phase
files’. These are well suited to coal since within a ramp area a set
of strips must normally be mined in order. Areas with different
ramp access often have several areas that can be advanced (or
delayed). The policies will control what actions to take in each
period in order to maximise the NPV as the resources are mined.
Spectrum Series Volume 17 111

A New Methodology for Flexible Mine Design
B Groeneveld1,2, E Topal3 and B Leenders4
ABSTRACT
Uncertainty and risk are invariably embedded in every mining project.
Mining companies endeavouring to maximise their return for
shareholders make important strategic decisions which take years or even
decades to ‘play out’. Therefore, developing a model that analyses the
potential payoff of a decision based on current fixed assumptions is
severely flawed. A model that incorporates uncertainty and is able to
adapt will help deliver a design with a better risk-return profile. In this
paper, a new approach is developed in order to have a design that is
flexible and able to adapt with change. This is achieved by developing a
mixed integer programming model that determines the optimal design for
simulated stochastic parameters. This research has incorporated
optionality (flexibility) in relation to mining, stockpiling, processing plant
and port capacity. The results are promising and are helping decision
makers to think in terms of value, risk and frequency of execution.
INTRODUCTION
Mining projects are characterised as being highly uncertain and
variable mainly due to the volatile nature of commodity prices
and uncertainty around geological conditions encountered in
orebodies. Uncertainty can arise from many different sources
including market prices, grade distribution, ground conditions,
equipment reliability, recovery of ore, human, capital and
legislative change (Topal, 2008). The mining industry will be
more sustainable if projects are developed in a manner that
increases the flexibility to respond to uncertainties in the
business cycle. For example, the global minerals industry has
seen an unprecedented demand for its products in recent years,
however the industry has struggled to change its level of supply
in response to price movements. Being able to design an
operation that has the flexibility to respond to this change
quickly will deliver better returns to stakeholders.
Geological uncertainty and risk have been incorporated in
optimum mine planning and design by a few studies to date.
Ramazan and Dimitrakopoulos (2004) developed a stochastic
based mixed integer programming (MIP) model for multiple
elements that uses several simulated orebodies in order to
minimise the grade uncertainty in the life of the mine schedule.
This model also takes into account risk quantification, equipment
access and mobility, and other operational requirement such as
blending, mill capacity and mine production capacity. Godoy and
Dimitrakopolus (2004) developed a new set of ways to generate a
mine production schedule under geological uncertainty. The first
stage of the method generates a stable solution domain which
shows the possible ore and waste extraction rates for a given
open pit. The second stage generates optimum ore production
and waste removal under uncertainty. The third stage generates a
series of physical schedules which obey slope constraints,
1. GAusIMM, PhD Candidate, Mining Engineering Department,
Western Australian School of Mines, Curtin University of Technology,
GPO Box U1987, Perth WA 6845.
Email: benjamin.groeneveld@gmail.com
2. Newcrest Graduate Mining Engineer, Australia.
3. MAusIMM, Associate Proffesor and Head of Mining Engineering
Department, Western Australian School of Mines, Curtin University of
Technology, GPO Box U1987, Perth WA 6845.
Email: e.topal@curtin.edu.au
4. MAusIMM, Principal Advisor – Strategic Development, Rio Tinto
Iron Ore, Expansion Projects, Resource Development Division, Perth
WA 6000. Email: bob.leenders@riotinto.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
maximise the equipment utilisation and meet mill requirements
while matching the mining rates previously derived by the
optimisation. The last stage generates a single mining sequence
from alternative sequences produced in the third stage by using a
new algorithm based on the simulated annealing method. Leite
and Dimitrakopoulos (2007) examined a stochastic optimisation
model based on simulated annealing algorithm for open pit
mines and applied it to a copper deposit for risk analysis. The
study shows the stochastic approach generates 26 per cent higher
net present value (NPV) than the conventional scheduler. Also,
the study suggests that life-of-mine schedules which incorporate
geological uncertainty lead to more informed investment decisions
and improved mining practices. Leite and Dimitrakopoulos (2010,
in this volume) and Ramazan and Dimtrakopoulos (2007) present
a stochastic integer programming (SIP) formulation for mine
production scheduling. The approach explicitly integrates the
grade uncertainty in the scheduling problem to generate a more
risk-robust solution. The same copper deposit has been utilised in
Leite and Dimitrakopoulos (2007) for application and the
stochastic formulation is shown to produce 29 per cent higher
NPV than the schedule obtained from conventional scheduling.
Meagher, Sabour and Dimitrakopoulos (2010, in this volume)
introduced an approach to integrate the block destination
flexibility in the process of assigning a value to mining blocks in
planning process via real options valuation (ROV) considering
geological and market uncertainties. The proposed approach;
firstly assigns a dollar value to each mining block considering
the different aspect of uncertainty and management flexibility.
Secondly, it utilises a minimum cut algorithm for designing a
lower risk long-term mine plan. Application to a case study
demonstrates significant differences in block value estimates.
A developing decision-making tool aimed at increasing the
flexibility of an engineering system is real options ‘in’ projects.
Significant research into this method has been undertaken by de
Neufvillie and his colleagues with applications in various
industries. This method is located midway between financial real
options analysis (which does not deal with system flexibility) and
traditional engineering approaches (which does not deal with
financial flexibility). A popular example used to explore the
concept of real options ‘in’ projects is that of a multi-story car
park. Flexibility in this situation is in the design of the footing and
columns of the building so that additional levels can be added at a
later date. This flexibility comes at a cost, and the designer must
determine if this is warranted. An example of the opportunity this
technique poses for mining is the Chilean mine in the ‘Cluster
Toki’ region, and was undertaken by de Neufville , in conjunction
with Codelco. In this example, a staged development of the real
option ‘in’ projects methodology is used where different operating
plans are designed to respond to changing prices. Truck fleet
capacity and crusher size were altered in the different operating
plans. The application of this method resulted in approximately 30
to 50 per cent more accurate project value than current estimates.
This approach provides a strong basis on which to grow real
options ‘in’ projects theory for mining. However, there are several
deficiencies in the current model. First, the initial scenario
construction used in the model limits the flexibility up front in the
model and prevents the optimal design from being chosen.
Therefore, how useful is this technique for valuing flexibility?
Secondly, the model fails to deal with variations in grade and
recovery in a transparent manner; one of the key drivers. Finally,
Spectrum Series Volume 17 113
B GROENEVELD, E TOPAL and B LEENDERS
the model does not incorporate options at all stages of a typical
mine value chain (de Neufville, Scholtes and Wang, 2005; Wang
and de Neufville, 2005, 2006; Cardin, 2007; Cardin, de Neufville
and Kazakidis, 2008). This paper outlines a new methodology to
evaluate the flexibility of strategic mine design under uncertainty,
using mixed integer programming (MIP) and Monte Carlo
simulation (MCS). An application of this methodology to a
hypothetical case study will be undertaken in order to show the
power of the model to handle complex strategic decisions.
METHODOLOGY
In order to evaluate the flexibility in strategic mine design, this
research employs MIP and MCS. In particular, MIP allows for ‘go’
or ‘no go’ decisions to be modelled for optimal execution under a
set of uncertainties. Uncertainties (or stochastic parameters) can be
simulated using MCS. In this way, each model (or trial) represents
a single path of a lattice tree (or binomial tree).
Description of model components
The model consists of three main components which feed the
MIP model; resource model, design options and stochastic
parameters (Figure 1). Running the model multiple times
generates a database of optimal designs for a given ‘state-of-the-
world’. This data set then provides a pathway to determine the
flexibilities that provide the best risk-return profile.
FIG 1 - Conceptual diagram of how the various model components
feed the mixed integer programming model and the result set.
FIG 2 - Material flow in the model and the location of the design options.
114 Spectrum Series Volume 17
Overview of resource model
Resource characteristics are a driving force in mine design. The
MIP model uses a resource model to provide a representation of
material that is available for processing through the life-of-mine
(both ore and waste is considered). The representation of the
resource is carried out by parcels of material. A parcel of
material can be defined as a quantity of material with an average
grade determined by the weighted average of the grade bins
contained within the parcel. A parcel may be made up of one or
more grade bins. A grade bin represents a quantity of material at
a specified grade. This is incorporated to provide a higher level
of detail to the model which will alter the decisions on how
material is processed, whilst minimising the number of integer
variables. These parcels are designed to represent a physical
constraint on the resource. The most common physical constraint
is the vertical mining constraint, which is included in the model
through parcel dependency. Mining of the grade bins within a
parcel can occur in any order, as long as the average parcel grade
(within a nominal deviation) is extracted each period. This forces
the model to take waste and ore in the same proportion.
Overview of the design options
Flexibility is included through various design options in the MIP
model. Solving the MIP models will determine which options are
executed and when. A full set of design options is dynamically
incorporated in the model which determines if and when these
options should be executed. These options are broken into four
categories: mine, preprocessing stockpiling, processing plants
and port capacity. More than one option type can be executed in
each period, hence these are not mutually exclusive decisions.
An illustration of the material flow and points where design
options may occur is shown in Figure 2. Some assumptions have
been made to simplify the model at this early stage of
development. These assumptions can be removed with further
refinement to the model; one type of circuit exists in each plant
with one set of beneficiation characteristics; port stockpiles are
not available in the model. This means the model must ship
material as soon as it is processed.
Available mine options in the model
Mine options are incorporated in the model to reflect mining
capacity constraints that exist in an operation. It is not feasible to
have unlimited mining capacity, due to the high capital cost
associated with additional capacity and/or technical pit constraints
(geotechnical and equipment interaction). Mine options can be
modelled to reflect truck capacity or shovel capacity. This type
of decision is repeatable many times in each period (ie you can
Advances in Orebody Modelling and Strategic Mine Planning I
 A NEW METHODOLOGY FOR FLEXIBLE MINE DESIGN

purchase more than one truck of the same type), thus mine
options are represented as integers. This allows for one or more
trucks of the same type to be purchased in each period.
Available stockpile options in the model
Stockpiling is used in mine operations for many reasons including
blending of material, storage of excess mine production and
storage of low-grade ore for future production. Long-term
stockpiling is included in the model, allowing material to be stored
on a stockpile in time (t) and removed in subsequent periods
(t+1…t+N). A further ability of the stockpile option is its ability to
represent long-term waste dumps. This functionality allows the
model to consider waste movement, and dynamically changes the
cut-off grade. Waste dumps are developed by entering an option
that is similar to a stockpile but has no plants for the material to
flow to, forcing it to remain on the waste dump.
Outline of what stochastic parameters are included
Uncertainty in the mining process is incorporated through the
market price, cost (capital and operating), utilisation of
equipment, plant recovery and time to build an option. Values for
these various inputs are simulated through a MCS process.
Model formulation
The developed MIP model optimises the available mine,
stockpile, plant and port flexibility for a simulated scenario.
These various design options dictate how the system is
configured and consequently the amount of production that can
occur. They also dictate the financial viability of the operation
and drive both revenue and operating costs. An outline of the
mathematical formulation is provided below.

Available plant options in the model

Objective function
Plant flexibility is incorporated to model the options managers
have around processing of mined ore through varying plant The objective function seeks to maximise before tax net present
designs. A plant option is characterised by its capacity, capital value (NPV):
cost, fixed operating cost, recovery characteristics and grade
T L M S L O
limits. A processing plant is the link between the mine and the 1
port in the flow of material in the model. Plant options can also be ∑ (1 + r ) [ S ∑ M
* t t l,t − ∑ C m I m , t − ∑ C sY s , t − ∑ C lYl , t − ∑ C oYo , t
t=1 l=1 m=1 s=1 l=1 o= 1
dependent on other plant options being built, allowing the idea of M L O M L

modular plant capacity to be modelled. That is where a high initial − ∑ Dm IDm , t − ∑ Dl IDl , t − ∑ Do IDo , t − ∑V m, t X m , t − ∑ Vl , t X l , t
m=1 l=1 o= 1 m=1 l=1
capital cost is incurred to allow for expansion at a later date with a
smaller capital cost.
S O M L O
− ∑ Vs , t X s , t − ∑ Vo , t X o , t − ∑ Fm , t I m , t − ∑ Fl , tYl , t − ∑ Fo , tYo , t
Available port options in the model s=1 o= 1 m=1 l=1 o= 1

Port options allow the sale of material to customers in the model.

M L O L P, B
Enough port capacity must exist in a period in order for any plant
production to occur in that period. For example, if we have two ∑ FR m, t IDm , t − ∑ FRl , t IDl , t − ∑ FRo , t IDo , t − ∑ L l ∑X p, b , l , t
m=1 l=1 o= 1 l=1 p= 1, b = 1
million tonnes of plant capacity and no port capacity in the first
period, then the production from the plant is forced to be zero. S P, B, L
−∑ L s ∑ XI ⎤
Overview of how mine scheduling works in the model s=1 p = 1 , b = 1 , l = 1 s ∈l
p,b,s,l,t ⎦⎥

Resource characteristics are a driving force in mine design. The

MIP model uses a resource model to provide a representation of where:
material that is available for processing through the life of mine
(both ore and waste is considered). r* is the rate of return on the project
The representation of the resource is carried out by parcels of St is sale price in time period t (in $/metal unit)
material. A parcel of material can be defined as a quantity of Ml,t is the metal units exiting plant l in time t
material with an average grade determined by the weighted
average of the grade bins contained within the parcel. A parcel Cm, Cs, Cl, Co is the capital cost of mine m or stockpile s or
may be made up of one or more grade bins. A grade bin represents plant l or port o
a quantity of material at a specified grade. This is incorporated to Im,t is the execution integer on mine option m in
provide a higher level of detail to the model which will alter the time t
decisions on how material is processed, whilst minimising the
number of integer variables. These parcels are designed to 0 is no execution of option;
represent a physical constraint on the resource. The most otherwise, is the number of times the option
common physical constraint is the vertical mining constraint is executed in time t.
which is included in the model through parcel dependency.
Ys,t,Yl,t,Yo,t is the execution binary on stockpile s or plant
Mining of the grade bins within a parcel can occur in any order l or port o
as long as the average parcel grade (within a nominal deviation)
is extracted each period. This forces the model to take waste and 0 is no execution of option;
ore in the same proportion. otherwise, the option is executed in time t.
Options in the model to test best plant or stockpile Dm, Ds, Dl, Do is the disposal cost of mine m or stockpile s
location through varying mining costs or plant l or port o

The optimal location for a processing plant varies with time as IDm,t is the disposal integer on mine option m in
the resource is mined in different regions. Therefore, determining time t
the best location for a plant or a stockpile is not a simple case 0 is no disposal occurs in time t;
and must consider these multiple uncertainties as it will most
otherwise, is the number of options disposed
likely change over the life of the project. Different plant
of in time t.
locations and stockpile locations can be tested by developing a
mining cost which varies by parcel and destination in the model. IDl,t ,IDo,t is the disposal integer on plant l or port o

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 115
B GROENEVELD, E TOPAL and B LEENDERS

B, t t
0 is no disposal occurs in time t;
∑ X p + 1 , b , tt ≤ R p + 1 ∗ ∑ Y p , tt ∀p, t
otherwise, is option is disposed in time t. b = 1 , tt = 1 tt = 1

Vm,t, Vs,t, Vl,t, Vo,t is the variable cost of mining a tonne of ore where:
from mine m or stockpile s or plant l or port Rp+1 is the resource of the successor parcel p + 1
o in time t
Xp+1,b,tt is the tonnage mined from the successor parcel p + 1 bin
Xm,t, Xs,t, Xl,t, Xo,t is the tonnage processed through mine m or b in time tt
stockpile s or plant l or port o in time t
Sequencing constraint 3
Fm,t,Fl,t,Fo,t is the fixed cost of mining from mine m or
plant l or port o in time t This constraint is a set packing constraint that forces a parcel to
only be fully mined once:
FRm,t,FRl,t,FRo,t is the reduction in fixed cost of disposing of
T
an option in time t
∑Y p, t ≤1 ∀p
Ll is the cost of mining a tonne of ore to plant l t=1

Ls is the cost of mining a tonne of ore to Equal mining of a parcel

stockpile s
This constraint forces the model to take high-grade ore, waste and
Xp,b,l,t is the tonnage mined from parcel p bin b to low-grade ore in equal proportions. This prevents the model taking
plant l in time t high-grade in the first period, followed by low-grade in the second
XIp,b,s,l,t is the tonnage mined from parcel p bin b to period and waste in the following period. A minimal deviation (γ)
stockpile s to plant l in time t of two per cent was allowed to prevent an infeasible solution:
The objective function represents the following: B
⎡ B ⎤
∑X p, b , t G p , b ≥ ⎢∑ X p , b , t ⎥G p ∗(1 − γ%) ∀p, t
The revenue from the sale of the ore less the b =1 ⎣b = 1 ⎦
capital cost of building an option less the B
⎡ B ⎤
disposal cost of reducing capacity less the
variable cost of processing ore less the fixed cost ∑X p, b , t G p , b ≤ ⎢∑ X p , b , t ⎥G p ∗(1 + γ%) ∀p, t
b =1 ⎣b = 1 ⎦
of maintaining an option; all multiplied by the
relevant discount factor for the cash flow in time t. where:
The model seeks to maximise this relationship. Gp,b is the grade of parcel p and bin b
The constraints in the model can be divided into five Gp is the grade of parcel p
categories: production, mining, stockpiling, processing plant and
port constraints. Flow balance constraint
This constraint links the flow paths in the model and ensures that
Production constraints the material available to the processing plant and stockpiling
options originates from the resource:
Resource constraint L S,K,L
X p, b , t = ∑ X p, b , l , t + ∑ XI p , b , s , k , l , t ∀p, b, t
This constraint makes sure the total amount of material extracted l=1 s = 1 , k = 1 , l = 1 s ∈l
from a mining pit has an upper bound based on the resource.
This constraint is applied at a parcel and bin level in the model: Mining constraints
T

∑X p, b , t −R p , b ≤ 0 ∀p, b Mining requirements

t=1
The constraint makes sure that mining includes all movement to
where: plant options, stockpile options and movement off stockpiles:
Xp,b,t is the tonnage mined from parcel p bin b in M P, B S,L
time t ∑X m, t = ∑ X p, b , t + ∑ XOs , l , t ∀t
m=1 p= 1, b = 1 s = 1 , l = 1 s ∈l
Rp,b is the resource of parcel p bin b
where:
Sequencing constraint 1
XOs,l,t is the tonnage sent from stockpile s to location l in time t
This constraint in conjunction with the next constraint forces the
binary value to be one in the period the parcel is fully mined. Mining capacity limit
This then allows the model to mine any successor parcels of ore:
B, t
This constraint ensures that mining only occurs if there is
∑ X p , b , tt ≥ R p ∗Y p , t ∀p, t sufficient capacity in a period to handle the movement. Capacity is
determined dynamically based on when mining options are
b = 1 , tt = 1
executed:
where: t t

Rp is the resource of parcel p ∑A I

m, u , t m, u − ∑ Am , u , t IDm , u ≤ X m , t ∀m, t
u =1 u =2

Sequencing constraint 2 where:

This constraint ensures that a parcel’s predecessor is mined Am,u,t is the capacity of mine option m that was executed in
before the successor is mined: period u in time t

116 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 A NEW METHODOLOGY FOR FLEXIBLE MINE DESIGN

Mine option disposal constraint Stockpile capacity constraint

This constraint ensures disposal of an option can only occur if This constraint makes sure the tonnage of material stockpiled
the option has been built. For example, if a mine option is built in across all plant locations does not exceed the stockpile capacity:
period one, and in period five there is no more material to mine,
L,t L,t t
then the model can dispose of this capacity in order to reduce the
fixed cost incurred: ∑ XI s , l , tt − ∑ XOs , l , tt ≤ ∑ A sY s , tt ∀s , t
l = 1 , tt = 1 s ∈l l = 1 , tt = 2 s ∈l tt = 1
t
IDm , t ≤ ∑ I m,tt − 1 ∀m, t where:
tt = 2
XOs,l,t is the total tonnage sent from stockpile s to plant l in
t t
time t
∑ ID m , tt ≤ ∑ I m,tt − 1 ∀m, t
As is the total capacity of stockpiles
tt = 2 tt = 2

Maximum execution Stockpile grade constraint on bins

Since the mine option can be modelled as an integer variable, This constraint applies the grade limits of the stockpile bins to
limits on how many times it can be applied in the model may be material entering each stockpile bin:
included. This is an optional constraint which can be turned on or t t
off when running the model. ∑ G p , b XI p , b , s , k , l , tt ≤ ∑ GU s , k XI p , b , s , k , l , tt ∀p, b, s, k, l, t
tt = 1 tt = 1

t t
Period constraint
∑ G p , b XI p , b , s , k , l , tt ≥ ∑ GL s , k XI p , b , s , k , l , tt ∀p, b, s, k, l, t
This constraint restricts the number of mine options built in a tt = 1 tt = 1
period to the period constraint maximum:
where:
I m,t ≤ PGm ∀m, t GUs,k is the upper grade limit of stockpile s bin k
GLs,k is the lower grade limit of stockpile s bin k
Overall constraint
This constraint restricts the total number of mine options built Bin removal constraint
over the life of a mine:
The constraint ensures material moved from the stockpile has
T been added to the stockpile at least one period ago, and that
∑I m, t ≤ OC m ∀m material removed from the stockpile is not removed again:
t=1
t t

where: XOs , k , l , t ≤ ∑ XI s , k , l , tt + ∑ XOs , k , l , tt − 1 ∀s, k, l, t

tt = 1 tt = 2
PC is the period constraint limit or the maximum number of
times an option can be executed in any period where:
OC is the overall constraint which is the maximum number of XOs,k,l,t is the tonnage removed from stockpile s bin k to plant l
times an option can be executed over the life of the project in time t

Stockpiling constraints Bin extraction

This constraint makes the total tonnage of material that is
Total inflow constraint extracted from each stockpile bin equal the overall extraction
from the stockpile:
This constraint makes the total amount of material that is entering K
a stockpile equal the material entering each stockpile bin: ∑ XOs , k , l , t = XOs , l , t ∀s, l, t s ∈ l
k= 1
K
XI s , l , t = ∑ XI s , k , l , t ∀s, l, t s ∈ l
k= 1 Metal extraction
where: This constraint ensures the total amount of metal units extracted
from a stockpile equals the metal units extracted from the
XIs,l,t is the total tonnage sent into stockpile s at plant l in time t
individual grade bins:
K
Flow balance constraint MOs , l , t = ∑ GA s , k XOs , k , l , t ∀s, l, t s ∈ l
k= 1
This constraint restricts the total amount of material coming into
each bin in the stockpile to be equal to the material sent from the where:
each parcel to the bin:
MOs,l,t is the metal units removed from stockpile s at location l
P, B
in time t
XI s , k , l , t = ∑ XI p , b , s , k , l , t ∀s, k, l, t s ∈ l
p= 1, b = 1 GAs,k is the average grade of stockpile s grade bin k
where: Opening limit
XIs,k,l,t is the tonnage sent into stockpile s grade bin k at
plant l in time t This constraint ensures that a stockpile can only be opened once:
T
XIp,b,s,k,l,t is the tonnage from parcel p bin b sent into stockpile ∑ Y s , t ≤ 1 ∀s
s grade bin k at plant l in time t t=1

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 117
B GROENEVELD, E TOPAL and B LEENDERS

S
Processing plant constraints
∑ E l , tGM l , t MOs , l , t ∀l, t
s = 1 s ∈l
Grade limits (upper and lower)
where:
This constraint applies the grade limits on a given plant in each
time period. This ensures every plant processes material it can GMl,t is the grade multiple of plant option l in time t
handle:
S P, B Plant option dependency
∑ MOs , l , t + ∑ G p, b X p, b , l , t
Plant option dependency dictates the relationships that occur
s = 1 s ∈l p= 1, b = 1

S P, B
between options. Two types of relationships are available;
≤ GUl ∑ XO s,l,t + GUl ∑ X p , b , l , t ∀l, t one-for-one and one-for-many.
s = 1 s ∈l p= 1, b = 1
One-for-one relationship
S P, B S

∑ MOs , l , t + ∑ G p , b X p , b , l , t ≥ GLl ∑ XO s,l,t This constraint makes sure a successor option is built prior to the
s = 1 s ∈l p= 1, b = 1 s = 1 s ∈l predecessor option being built in an equal ratio. For example,
P, B this can be used to model a modular plant design where an initial
+ GLl ∑ X p, b , l , t ∀l, t investment can be made in plant capacity that has the ability to
be expanded easily for a lower capital than if the initial investment
p= 1, b = 1
was not made (this later expansion is optional):
where: t t − DT

GLl is the lower grade limit of plant l ∑Y l , tt ≤ ∑ Yc , tt ∀l, c, t c ∈ l

tt = 1 tt = 1
GUl is the upper grade limit of plant l
One-for-many relationships
Plant capacity constraint This constraint allows a successor option to be built if its
The constraint ensures that the total tonnage of material processed predecessor option has been built at least once. This can be used
in a period shall be less than the capacity of plant options built and to model a rail link to a plant location where an initial capital
disposed: investment is required. However, once this has been built,
S P, B t t
numerous plants can be built at the same location:
∑ XOs , l , t + ∑ X p , b , l , t ≤ ∑ Al , u , tYl , u − ∑ Al , u , t IDl , u ∀l, t t − DT
s = 1 s ∈l p= 1, b = 1 u =1 u =2 Yl , t ≤ ∑ Yc , tt ∀l, c, t c ∈ l
tt = 1
where:
where:
Al,u,t is the capacity of plant option l built in u in time t
c is the predecessor plant option of plant option l
Plant disposal constraint DT is the lead time on the relationship

This constraint ensures that a plant option is only disposed if the Yc,tt is the execution variable of the predecessor plant c of plant
plant has been built in a previous period. This will result in a l in time tt
fixed cost saving, however an additional disposal cost will be
incurred in the objective function: Port constraints
t
YDl , t − ∑ Yl , t ≤ 0 ∀l, t Port production constraint
tt = 1
The constraint ensures the total tonnage of material processed
t t
through all plant options is less than or equal to the total port
∑ YD l , tt − ∑ Yl , tt ≤ 0 ∀l, t
capacity:
tt = 1 tt = 1
L O

Tonnage produced ∑X l,t ≤ ∑ X o, t ∀t

l=1 o= 1

This constraint restricts the tonnage exiting the plant to be equal

to the material entering the plant multiplied by the plant recovery: Capacity constraint
P, B S
The constraint requires the total tonnage of material shipped in a
Xl , t = ∑ E l , t X p, b , l , t + ∑ E l , t XOs , l , t ∀l, t period to be less than the port capacity:
p= 1, b = 1 s = 1 s ∈l
t t
where: X o , t ≤ ∑ Ao , u , tYo , u − ∑ Ao , u , t IDo , u ∀o, t
u =1 u =2
El,t is the recovery of plant l in time t
where:
Metal units produced Ao,u,t is the capacity of port option o built in u time t
This constraint calculates the metal production of a plant option by
multiplying the metal units in to the plant by the recovery and
Disposal constraint 1
grade multiples for the plant. This is used to calculate the revenue Disposal of a port option may only occur if the option has
of the mine: previously been built:
P, B t
Ml,t = ∑E l,t GM l , t X p , b , l , t + IDo , t ≤ ∑ Yo , tt ∀o, t
p= 1, b = 1 tt = 1

118 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 A NEW METHODOLOGY FOR FLEXIBLE MINE DESIGN

t t
haul for pit 1, and location B has the shortest haul for pit 3,
∑ ID o, t ≤ ∑ Yo , tt ∀o,t
whilst pit 2 has an equivalent haul to either location). In order to
tt = 1 tt = 1
process material through a plant at location A, a rail link of 35km
with a capital of $65 M needs to be built. Likewise, at location B,
Port option dependency
a rail link of 10km needs to be built for a capital of $30 M.
Port option dependency may occur in one-for-one or one-for- Finally, in order to process any material through the port, a rail
many relationships. link from the junction of A and B to the coast needs to be built
for a capital of $20 M. The analysis will look at the system
One-for-one relationship configuration over five periods.
This constraint makes sure a successor option is built prior to the
predecessor option being built in an equal ratio:
t t − DT

∑Y o , tt ≤ ∑Y c , tt ∀o, t c ∈ o
tt = 1 tt = 1

One-for-many relationships
This constraint allows a successor option to be built if its
predecessor option has been built at least once. This relationship
can be used to model a rail link that must be built before any port
can be built:
t − DT
Yo , t ≤ ∑Y c , tt ∀o, t c ∈ o
tt = 1

where:
DT is the lead time on the relationship
Yc,tt is the execution variable of the predecessor port c of port o FIG 3 - Conceptual layout of hypothetical mine.
in time tt

Non-negativity, binary and integer restrictions The model inputs

Multiple options were included in the model of this problem as
Non-negativity summarised in Table 1. A full list of the fixed costs, variable
costs and grade constraints is not provided for simplicity
The following variables are restricted to taking on positive values purposes. In order to simulate the different processing plant
as a negative would represent an infeasible situation: locations available to the model, a differential mining cost was
X p , b , t , XI p , b , s , k , l , t ≥ 0 ∀p, b, s, k, l, t used based on the destination of material. A summary of the
different costs associated with each location is outlined in Table 2.
X m , t , X l , t , X o , t ≥ 0 ∀m, l, o, t
XOs , l , t , XOs , k , l , t , XI s , k , l , t XI s , l , t ≥ 0 ∀s, k, l, t TABLE 1
Options available in hypothetical example.
Integers
Type Cost ($M)
The following variables must take on integer values in the model:
Mine options
I m , t , IDo , t , IDm , t , IDl , t ∀m, l, o, t 1 Mt/a unit 3
2 Mt/a unit 4.5
Binaries
Stockpile options
The following variables must take on binary values; integers with
Waste stockpile (500 Mt capacity) 0
an upper bound of one and lower bound of zero:
Low-grade (30 Mt capacity) 0
Yc , tt , Y p , t , Yo , t , Yl , t , Y s , t ∀p, o, l, s, t, c, tt
Plant options
CASE STUDY – OPEN PIT MINE 5 Mt/a fixed (lead 0.5 yr) 50
10 Mt/a fixed (lead 1.5 yr) 92
An application of the methodology was implemented to a
hypothetical mining scenario. The problem is similar in nature to 5 Mt/a flexible (modular, lead 0.5 yr) 75
an iron ore mine, although it could be applied to any open cut Additional 5 Mt/a flexible (modular, lead 0.5 yr) 30
mine. A single mine site is used in this example. Port capacity
10 Mt/a (lead 0.75 yr) 100
The problem 20 Mt/a (lead 1.5 yr) 175
The operation consists of three mining pits (two high-grade and
one low-grade), two plant locations with associated rail
infrastructure, stockpiling and waste storage capabilities at each The stochastic variables
location, and two port options with associated rail requirements In this problem, it was determined that seven stochastic parameters
(Figure 3). In this diagram, the rectangular boxes represent would be included. These were price, recovery, capital cost,
different plant locations (note that location A has the shortest operating cost, and utilisation for mine, plant and port options.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 119
B GROENEVELD, E TOPAL and B LEENDERS

TABLE 2 Results analysis

Differential mining costs to handle different locations in the model.
Based on the input parameters, 200 trials were run, with
Cost ($/t) Pit 1 Pit 2 Pit 3 CPLEX™ used to solve the MIP model. In total it took three
Plant location A 0.2 0.6 1.8 hours to process the model, which was deemed a good solution
Plant location B 2.2 0.6 0.4 time for this model size. The raw data from the results exceeds
four gigabytes. A results analysis process has been developed
Waste location A 0.1 0.3 0.9
which summarises this data. After processing of the model, the
Waste location B 1.1 0.3 0.2 frequency of execution for each options was analysed (Table 3).
Frequency of execution is calculated by dividing the count of the
Choice of underlying distributions was done through discussions number of times an option is executed by the maximum number
with professionals. No detailed analysis of the underlying nature of times it could be executed. Some categories of options (plant,
of the stochastic variables has been carried out, as detailed port, mine) sum to more than one hundred per cent because
research in other papers is available, which was not the primary multiple expansions of that type can occur in the same time
purpose of this paper (Dimitrakopoulos and Abdel Sabour, 2007; period as the options are not mutually exclusive.
Godoy and Dimitrakopoulos, 2004; Lima and Suslick, 2006; From examining Table 3, several conclusions can be
Morley, Snowden and Day, 1999; Topal, 2008). developed. First, it is evident that larger port capacity options
A summary of the values used for each distribution is as should be investigated as the execution frequency is over 50 per
follows: cent for the 20 Mt/a port option (the largest in the model) for
• price follows a lognormal distribution with a mean of $85, three periods. Second, the mine option with 1 Mt/a capacity and
standard deviation of $25 and a correlation of 0.30 between the plant with 5 Mt/a at location B are not valuable options, as
periods; their execution is lower than 20 per cent. Finally, location A is
• recovery follows a triangular distribution with a maximum preferred over location B, as the rail link options which dictate
value of 90 per cent, likely value of 80 per cent, minimum which locations can be used are executed 94 per cent and 45 per
value of 70 per cent and a correlation of 0.05 between periods; cent for A and B respectively. A value at risk graph (VARG)
shows the risk to return relationship. Figure 4 displays the VARG
• capital cost multiple follows a normal distribution with a
mean of 1.08, standard deviation of 0.20 and a correlation of
0.40 between periods;
• operating cost multiple follows a normal distribution with a
mean of 1.03, standard deviation of 0.10 and a correlation of
0.10 between periods;
• mine equipment utilisation follows a triangular distribution
with a maximum value of 95 per cent, likely value of 75 per
cent, minimum value of 60 per cent and a correlation of 0.22
between periods;
• plant utilisation follows a triangular distribution with a
maximum value of 95 per cent, likely value of 80 per cent,
minimum value of 65 per cent and a correlation of 0.21
between periods; and
• port utilisation follows a triangular distribution with a
maximum value of 95 per cent, likely value of 80 per cent,
minimum value of 65 per cent and a correlation of 0.34
between periods. FIG 4 - Value at Risk for case study.

TABLE 3
Frequency of execution for all options in the model.
Period 1 (%) Period 2 (%) Period 3 (%) Period 4 (%) Period 5 (%) All periods (%)
Mine 1 Mt/a 18 4 5 0 0 5
Mine 2 Mt/a 98 32 9 3 1 29
Rail link to A 94 0 0 0 0 19
Rail link to B 45 1 0 0 0 9
Plant 5 Mt/a (A) 69 26 14 6 2 23
Plant 10 Mt/a (A) 66 14 5 1 0 17
Plant 5 Mt/a modular (A) 90 41 13 2 0 29
Additional 5Mt/a modular (A) 0 82 35 16 1 27
Plant 5 Mt/a (B) 19 4 1 1 0 5
Plant 10 Mt/a (B) 44 19 6 2 0 14
Plant 5 Mt/a modular (B) 41 15 4 2 0 12
Additional 5Mt/a modular (B) 0 35 14 3 2 11
Port 10 Mt/a 98 1 0 0 0 2
Port 20 Mt/a 91 84 53 28 8 53

120 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I

FIG 5 - Example of decision tree for a different case study.
for this example with the base case representing a fixed mine
design with no optionality. The design chosen for the base case
was based on the 50th percentile design when the model with
optionality was run. This design was then fixed in the MIP model
and reprocessed with the same uncertainties. This shows the
outcome of management not changing the operating policy of the
mine. The mean NPV of the base case was $702 M, and for the
case with options was $1298 M, an 85 per cent increase. Further
to these analysis methods, experimentation is currently underway
using various data mining techniques. An open source software
package called Rapid Miner is currently being used. An example
(from a different problem set) of the output generated is shown in
Figure 5. This example shows a decision tree with the associated
percentages of times the decision paid off highlighted in the
Yes/No boxes at the bottom of the nodes.
CONCLUSIONS
In conclusion, this paper has developed a methodology to evaluate
the strategic mine design flexibility under a stochastic
environment. The proposed methodology is a unique approach that
allows flexible mine designs to be justified. The decision-maker is
supported in their choice and refinement of mine design.
Increasing flexibility in mine designs would be advantageous for
responding to changing business conditions across the full
economic cycle. For the sake of comparison, the proposed
methodology has been implemented to a hypothetical mining
scenario. The results demonstrated that the value of expected NPV
increases by 85 per cent with flexible mine design, compared to a
design without flexibility. The paper illustrates how to incorporate
design options (flexibility) into a strategic mine plan in a manner
that proactively manages inevitable uncertainties. It is hoped this
research will help in justifying more flexible mine designs and
further the sustainability of the industry.
RECOMMENDATIONS
Whilst the model handles a simple case, currently further
research and model improvements continue in the following
areas:
• More detailed modelling which considers multiple process
options and multi-product options.
Advances in Orebody Modelling and Strategic Mine Planning I
A NEW METHODOLOGY FOR FLEXIBLE MINE DESIGN
• Handling of grade variability through the use of conditional
simulation methods will greatly improve the power of the
model (Ramazan and Dimitrakopoulos, 2004).
• Further investigation into appropriate results analysis
techniques is required to fully understand how the primary
question of flexibility is answered.
• MIP performance improvement algorithms need to be
investigated; these methods may include reducing the
feasible region with additional constraints, and/or developing
a node selection routine for the branch and bound algorithm
that exploits some of the nuances in the model.
• Application of this technique to underground mining is
needed to fully capture the options available to mine
management. In particular, incorporating the process to
optimise the open cut and underground transition point
would be highly beneficial. This would assist in strategic
planning for the entire orebody.
ACKNOWLEDGEMENTS
The authors wish to thank Rio Tinto Iron Ore for their support of
this project.
REFERENCES
Cardin, M A, 2007. Facing reality: Design and management of flexible
engineering systems, master’s thesis (unpublished), Engineering
System Divisions, Massachusetts Institute of Technology.
Cardin, M A, de Neufville, R and Kazakidis, V, 2008. Process to
improve expected value of mining operations, Transactions of the
Institutions of Mining and Metallurgy, Mining Technology, 117(2):
A65-A70.
de Neufville, R, Scholtes, S and Wang, T, 2005. Real options by
spreadsheet: Parking garage case example, Journal of Infrastructure
Systems, 12(2):107-111.
Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans
under uncertainty: Can the real options make a difference?,
Resources Policy, 32(3): 116-125.
Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty-based production
scheduling in open pit mining, SME Transactions, 316, 106-112.
Godoy, M and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling, SME Transactions, 316,
43-50.
Spectrum Series Volume 17 121
B GROENEVELD, E TOPAL and B LEENDERS
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimization model
for open pit mine planning: Application and risk analysis at a copper
deposit, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 116(3):A109-A118.
Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under
metal uncertainty – Application of stochastic mathematical
programming at an open pit copper mine and comparison to
conventional scheduling, in Advances in Orebody Modelling and
Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Lima, G A C and Suslick, S B, 2006. Estimating the volatility of mining
projects considering price and operating cost uncertainties,
Resources Policy, 3:86-94.
Meagher, C, Abdel Sabour, S A and Dimitrakopoulos, R, 2010. Pushback
design of open pit mines under geological and market uncertainties,
in Advances in Orebody Modelling and Strategic Mine Planning I
(ed: R Dimitrakopoulos), pp 291-298 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Morley, C, Snowden, V and Day, D, 1999. Financial impact of
resource/reserve uncertainty, The Journal of The Southern African
Institute of Mining and Metallurgy, 6: 293-302.
122 Spectrum Series Volume 17
Ramazan, S and Dimitrakopoulos, R, 2004. Traditional and new MIP models
for production scheduling with in-situ grade variability, International
Journal of Mining, Reclamation and Environment, 18(2):85-98.
Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
long-term production scheduling for open pit mines with a new
integer programming formulation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 385-392 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Topal, E, 2008. Evaluation of a mining project using discounted cash
flow analysis, decision tree analysis, Monte Carlo simulation and
real options using an example, International Journal of Mining and
Mineral Engineering, 1(1):62-76.
Wang, T and de Neufville, R, 2005. Real Options ‘in’ Projects, paper
presented to Ninth Real Options Annual International Conference
Paris, France.
Wang, T and de Neufville, R, 2006. Identification of real options ‘in’
projects, paper presented to 16th Annual International Symposium of
the International Council on Systems Engineering (INCOSE),
Orlando, US.
Advances in Orebody Modelling and Strategic Mine Planning I

Pit Optimisation and Mine Scheduling at the Montes Claros
De Goiás Nickel Laterite Deposit, Brazil
B Y Tomaselli1, E A Silva2, G J de Carvalho Pereira3, J A Hilário4, M C Botelho5 and
P C Rodriguez6
ABSTRACT
The nickel lateritic deposit of Montes Claros de Goiás (MCG) is located
in the west portion of Goiás state in Brazil. The deposit presents two
main ore types: limonite and saprolite. The MCG deposit is located on a
flat topography and the alteration profile for the deposit presents the
limonite over the saprolite. Limonite presents a lower MgO and Ni grade
and a higher Co grade when compared with the saprolite. A conceptual
analysis was carried out in order to obtain the optimised nickel recovery
through a study comparing two different technologies: High Pressure
Acid Leaching (HPAL) and Ferronickel (FeNi). The methodology used
was to generate a series of pit optimisations and subsequent strategic
mining scheduling, including capital expenditures consideration and cash
flow analysis. Considering that the correct selection of the optimum
ultimate pit is essential for the project profitability, the presented paper
shows that it is suitable in the early definition phase of the project, mainly
during prefeasibility studies.
INTRODUCTION
The nickel lateritic deposit of Montes Claros de Goiás (MCG) is
owned by Votorantim Metais (VM) and is located in the west
portion of Goiás state in the city of Montes Claros de Goiás,
which is 280 km away from Goiânia, GO, Brazil. Nowadays it has
more than 60 000 m of RC drilling holes, bench and pilot
metallurgical tests and is on a prefeasibility stage. Figure 1 shows
the location of the deposit. The total measured and indicated
resources estimated are 135.6 Mt at 0.83 per cent Ni and have the
following distribution for each rock type: limonite are 101.4 Mt at
0.72 per cent Ni, 3.6 per cent MgO, and 47.1 per cent Fe2O3, and
saprolite are 34.2 Mt at 1.15 per cent Ni, 18.6 per cent MgO and
21.2 per cent Fe2O3. The choice of the technology route and the
final product are directly dependent of the ore characteristics that
will feed the plant, mainly its chemical and mineralogical
characteristics. The lateritic deposit composition will be dependent
on the type of parent rock, the climate where the deposit was
formed and the weathering history. There are two main routes for
1. Mining Engineer, Votorantim Metais Ltda, Dept° Geologia/
Exploração Mineral, Edifício Esplanada, Praça Ramos de Azevedo,
206-15° andar – Conjuntos, 1530/1540 Centro, São Paulo 01037-010,
Brazil. Email: bruno.tomaselli@vmetais.com.br
2. Mining Engineer, Geoexplore Consultoria e Serviços, Av Afonso
Pena, 3294 conjunto 206, Cruzeiro, Belo Horizonte MG 30130-009,
Brazil. Email: eduardo.silva@geoexplore.com.br
3. Mining Engineer, Votorantim Metais Ltda, Dept°
Geologia/Exploração Mineral, Edifício Esplanada, Praça Ramos de
Azevedo, 206-15° andar – Conjuntos, 1530/1540 Centro, São Paulo
SP CEP 01037-010, Brazil. Email: guilherme.jose@vmetais.com.br
4. Mining Engineer, Geoexplore Consultoria e Serviços, Av Afonso
Pena, 3294 conjunto 206, Cruzeiro, Belo Horizonte MG 30130-009,
Brazil. Email: joao.hilario@geoexplore.com.br
5. General Manager Mineral Development, Votorantim Metais Ltda,
Dept°, Geologia /Exploração Mineral, Edifício Esplanada, Praça
Ramos de Azevedo, 206-15° andar, Conjuntos 1530/1540 Centro,
São Paulo 01037-010, Brazil.
Email: maurilio.botelho@vmetais.com.br
6. Mining Engineering Manager, Geoexplore Consultoria e Serviços,
Av Afonso Pena, 3294 conjunto 206, Cruzeiro, Belo Horizonte MG
30130-009, Brazil. Email: porfirio.cabaleiro@geoexplore.com.br
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
nickel lateritic deposits: pyrometallurgy and hydrometallurgy
(Davi, Bacon and Osborne, 2004). The choice of the best route
will be in accordance to the ore magnesia content and its nickel to
iron ratio. The degree of inherent homogeneity within a deposit
will also influence on this decision.
Typically, a study is conducted considering the best process
route to treat the ore presented in the reserve. This approach
usually considers only stand alone operations and will leave
behind any marginal ore that could be treated by other
technologies. Nowadays some operations that have conducted
their projects to stand alone options are considering process the
marginal ore with another technology in order to obtain a
maximum recovery of its deposit.
Considering these two methods for treating a nickel lateritic
deposit, some characteristics based on operations and projects
should be noted (Brook Hunt, 2007). Ferronickel is a high energy
consumption that requires hydroelectric energy for feasibility. It is
only applied for saprolite ores and it has 75 to 90 per cent of
nickel recovery and no cobalt recovery. The CAPEX of ferronickel
is in the order of US$28 - 40 thousands per ton of nickel produced.
HPAL makes large use of reagents and is very sensitive to sulfuric
acid consumption as it usually represents its main reagent
consumption and may compromise the project feasibility. Usually
HPAL is only applied to limonite ores with low magnesium
content. It consumes less energy and has a significant energy
cogeneration on the acid plant. It has more than 90 per cent of
nickel recovery and 85 to 90 per cent of cobalt recovery and
flexible product types. The CAPEX of HPAL is in the order of
US$45 - 60 thousands per ton of nickel produced. If we consider a
HPAL stand alone operation at MCG deposit we must consider
that the recommended Mg grade is six per cent, as values higher
than this will lead to a unprofitable acid consumption and also
scaling problems on the autoclaves. Saprolite usually cannot be
used on a HPAL operation as it has high magnesia contents,
however, using only limonite and discarding the saprolite is not
recommended as higher Ni grades are on saprolite ore. On the
other hand, if we consider a FeNi stand alone operation we must
consider that limonite is not suitable to be used due to its high iron
content and low nickel grade and a stand alone operation is
profitable but also too restricted because of the low mass available
of saprolite when compared to limonite. A complete pit
optimisation study was developed considering both stand alone
scenarios and also a hybrid operation to find out which one is the
best option. The complete list of the nine scenarios is presented on
Table 1.
METHODOLOGY
The methodology consists of:
• import the geological resource model and apply the dilution
based on a benchmarking on a similar operation;
• define the benefit function based on technical and economical
constraints and assign a process route for each rock type;
• select three notable points among the obtained pit outlines;
best Net Present Value (NPV), maximum metal recovered or
largest pit, and a location considered important for project
profitability, 20 years life-of-mine (LOM) operation for
instance; and
Spectrum Series Volume 17 123
B Y TOMASELLI et al

FIG 1 - MCG deposit location.

124 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 PIT OPTIMISATION AND MINE SCHEDULING AT THE MONTES CLAROS DE GOIÁS NICKEL LATERITE DEPOSIT, BRAZIL

resource and should not be located on drainages or access

TABLE 1 routes. The bench height selected for this study was 3 m and
Pit optimisation scenarios. the berm width was 4 m resulting on an overall slope angle
of approximately 40°. Figure 2 shows a schematic graphic
Scenario Route Shell representation of the mine considering the predicted destination
1a HPAL Best NPV for each rock type.
1b HPAL 20 years LOM Pit optimisation was performed using Whittle Four-X® to
obtain the benefit of each block (Randall, 1994). A block is
1c HPAL Maximum resource
considered profitable if the income that could be generated is
2a FeNi Best NPV higher than the cost of processing and mining the block and all
2b FeNi 20 years LOM the waste over it (Kennedy, 1990). Because of that, to calculate
2c FeNi Maximum resource
the profitability of a block it is necessary to define all the costs
related and the correlated income that this block could generate.
3a Hybrid Best NPV Process costs included all the supplies, energy, labour, freights
3b Hybrid 20 years LOM and G&A from the crusher till, the client excluding mine
3c Hybrid Maximum resource operation. Mine operation considered all the operations that are
necessary to deliver the ore into the crusher, including all the
auxiliary services. The assumptions that were made are described
• schedule and generate a discounted cash flow (DCF) to below:
compare these scenarios, including the capital Expenditure Costs:
(CAPEX) (Noronha and Gripp, 2001). • FeNi process: flat rate;
• HPAL process: flat rate + acid cost;
GEOLOGICAL MODELLING AND DILUTION
• acid cost: F(Ni, Mg, Al, Fe) × acid price;
The geological modelling for MCG deposit was internally
developed by the Votorantim Metais (VM) geology team. The • mine: loading + transport + auxiliary equipments cost; and
geological interpretation was based on the following chemical • transport cost was considered block-by-block and calculated
criteria: in accordance to the distance of this block to the crusher and
• soil (SO): Ni < 0.3%; Fe ≥ 15%; MgO < 4%; Al2O3 ≥ 10%; adjusted to the bench height of the block.
• limonite (OX): Fe ≥ 22%; MgO < 10%; Fe + Al2O3 ≥ 25%; Income:
• saprolite (SA): Fe < 22%; MgO ≥ 10%; • Ni and Co price: flat long term price,
• bedrock (BD): Ni < 0.6%; MgO > 10; and • FeNi Ni recovery was 90 per cent,
• chalcedony (CD): SiO2 > 70%. • HPAL Ni recovery was 91 per cent, and
The density used on the model was based on an average • HPAL Co recovery was 89 per cent.
obtained by density tests. A diluted model was created based on For HPAL there is also the possibility of upgrading the grades
experience with similar operations. The methodology assumed using a very simple circuit of beneficiation prior to the autoclave
material loss of the upper layer in contact with soil or outcrop, which was considered on this study.
and for the contact zone between saprolite and limonite it was For the determination of the technological cut-off grade, a
considered a dilution of the upper layer with the bottom layer. graphic varying the cut-off Ni grade on the ore was generated.
The dilution was only applied to the blocks that are in the contact The aim was to achieve an average Ni grade of 1.1 per cent on
zone. HPAL and 1.4 per cent on FeNi. The graphs are presented in
Figure 3.
PIT OPTIMISATION Considering the process constraints and the graphs presented,
As it was defined by the geologists, there are two main ore types a cut-off Ni grade of 0.77 per cent was used for limonite with
within the resource: limonite and saprolite. The first one was upgrade. For the saprolite a cut-off Ni grade of 1.0 per cent was
assigned to be processed by HPAL and the second one by FeNi. used. These values were defined as a technological cut-off and
All the other rock types are defined as waste. To be considered a determined by the process area as a target to obtain reasonable
reserve, the block must be classified as measured or indicated results for the plant. All this data was imported into the Whittle

FIG 2 - Schematic representation of the mine operation.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 125
B Y TOMASELLI et al

FIG 3 - Limonite and saprolite grade tonnage curves.

30 k US$/t Ni and the ramp up period: 65 per cent of capacity for

100%
first year. Also considered was the HPAL throughput rate of
90%
1.66 Mt/a, the Mg of six per cent maximum, the CAPEX:
80%
50 k US$/t Ni and the Ramp up period of 25 per cent of capacity
70%
for first year. Finally, the LOM: 20 years of operation (minimum)
60%
was considered. Based on the schedules, a complete discounted
NPV (%)

50%
40%
30%
20% HPAL FeNi
10%
0%
0% 10% 20% 30% 40% 50% 60% 70%
Mined ore (%)
FIG 4 - Optimal pit ore tonnage versus NPV considering different
technologies.
Four-X® software to run the optimisation. A graphic of all the
optimised pits was generated using the results obtained by pit
optimisation from the software. From this graph (Figure 4) three
notable points were distinguished: best NPV, maximum resource,
and 20 year LOM. Note that values are regularised to a 100 per
cent basis for the greatest value obtained. The data for these
points are summarised in Table 2. It is important to note that
HPAL grades for this table do not include the upgrade factor and
ore mass values are on a 100 per cent basis for the greatest value
obtained.
Pit optimisation results show that a stand alone HPAL operation
leaves behind as waste an important amount of high grade ore that
would probably improve the project NPV. On the other hand,
using only saprolite ore on a FeNi stand alone operation will mine
limonite as waste and dump the bigger part of the deposit as
waste. Despite the lower Ni grade of the hybrid scenario when
compared to the FeNi operation, it has the highest nickel recovery
and the lowest strip ratio compared to stand alone options.
MINE SCHEDULING
To complete the comparison between the scenarios, a strategic
mine scheduling was developed considering the following
constraints and assumptions, including the optimisation of Ni
grades for the first years and the start-up on second quarter of the
first year. Also considered were the FeNi throughput rate of
0.88 million tonnes per annum (Mt/a), the SiO2/MgO ratio of two
(maximum), the Fe of 15 per cent maximum, the CAPEX of
cash flow analysis (DCF) was carried out in order to generate the
economic drivers for the different scenarios of the project.
Indicators of value (NPV), return (IRR) and investment
Hybrid efficiency (PVR) were then generated. NPV (Net Present Value)
was obtained for the project considering the opportunity cost of
80% 90% 100%
capital for the business and all costs associated. IRR (internal
rate or return) is the project rate of growth it is expected to
generate, ie the value on which the NPV of the project is equal to
zero. PVR (Present Value Ratio) was used to evaluate the
efficiency of an investment and can be translated as dollars
generated (NPV) per dollar invested in terms of maximum
capital exposure. As there are two different technologies with
different capital costs it is important to make use of an
economical measure to assist in the best scenario evaluation. The
final analysis will lead to three possible scenarios: best HPAL
stand alone, best FeNi stand alone and best hybrid operation.
Table 3 shows results for all the scenarios. Values for Ni
production, NPV and PVR were regularised to a 100 per cent
basis considering the greatest value as 100 per cent.
Considering that these are mutually exclusive projects, the
NPV is the appropriate economical tool to compare them
(Stermole and Stermole, 2006). Use of IRR or PVR should be
done in an incremental analysis to choose the best alternative. In
accordance to Tables 2 and 3 the best scenarios are HPAL: 1a,
FeNi: 2a and Hybrid: 3a.
To check the NPV results an incremental analysis between the
best scenarios of each alternative was performed. The incremental
analysis was performed between the hybrid scenarios against each
one of the stand alone. Results showed that HPAL stand alone is
the worst option. Also, despite FeNi stand alone having a greater
IRR (Table 3), the incremental analysis showed that the hybrid
operation presents both positives NPV and PVR. If we consider a
budget constraint this analysis may help to decide the incremental
gain of each scenario against other options. Considering that there
is no budget constraint, the hybrid scenario presented the best
performance when compared to the others. Although this will be
considered the base case for the project, the other two scenarios
will also be studied as a trade-off. The next step was to detail the
mine scheduling for these scenarios considering all the constraints
and also addressing production issues that would make a suitable

126 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 PIT OPTIMISATION AND MINE SCHEDULING AT THE MONTES CLAROS DE GOIÁS NICKEL LATERITE DEPOSIT, BRAZIL

TABLE 2
Pit optimisation results for all scenarios (grades in percentage).

Scenario Ore % Ni Co MgO Fe2O3 SiO2 Al2O3 Strip ratio

1a 61.7 1.01 0.066 4.61 47.68 18.23 7.66 2.4
1b 53.4 1.04 0.068 4.50 48.37 17.73 7.52 2.1
1c 64.9 1.00 0.065 4.66 47.37 18.49 7.68 2.8
2a 31.6 1.40 0.030 20.08 21.15 36.11 3.20 2.9
2b 29.3 1.42 0.030 20.15 21.11 36.03 3.19 2.7
2c 34.9 1.37 0.030 19.93 21.33 36.17 3.28 3.5
3a 93.5 1.15 0.055 9.41 39.54 23.69 6.28 1.9
3b 73.4 1.17 0.058 8.86 40.97 22.61 6.35 1.6
3c 100.0 1.13 0.053 9.99 38.27 24.67 6.15 2.1

TABLE 3 global optimum. It is a good project practice that the earlier this
is defined in the project, the less costly it will be to the project in
Mine scheduling economical measures results.
the future. The solution was to assign each rock type to the
Scenario LOM Production NPV IRR PVR process route that it is more suitable. This will ensure that no
years Ni % % % % additional cost will be expended to process that ore and also
1a 22 56 53.8 19.9 20.7
guarantee an optimum reserve. The MCG deposit is very unique,
saprolite has the greatest nickel grades but the limonite mass is
1b 21 56 52.7 20.1 32.1 too high to simply discard. Because of that, a hybrid operation is
1c 23 56 53.6 19.5 32.6 more profitable than just a stand alone. The method is also
2a 22 44 44.1 26.0 49.9 important because sometimes a stand alone operation is not
profitable, but when combined in a hybrid operation the synergy
2b 21 44 43.9 26.4 49.5 between the two could increase profitability. If we consider that
2c 25 44 43.4 25.3 49.0 the capital cost would decrease because of some synergies, the
3a 29 100 100.0 23.8 100.0 difference between the hybrid and the stand alone scenarios
would be greater.
3b 21 100 90.2 23.1 90.1
3c 31 100 83.3 19.0 83.3
REFERENCES
Dalvi, A D, Bacon, W G and Osborne, R C, 2004. The past and the future
and more realistic mine scheduling in terms of operation. That of nickel laterites [online], PDAC 2004 International Convention,
work is not part of this paper. pp 1-27. Available from: <http://www.pdac.ca/pdac/publications/
papers/2004/techprgm-dalvi-bacon.pdf> [Accessed: 29 September
2008].
CONCLUSIONS Kennedy, B A, 1990. Surface Mining, 1206 p (Society for Mining,
Metallurgy and Exploration: Littleton).
It is well known that all lateritic deposits usually present
Noronha, R A and Gripp, A H, 2001. Ultimate pit selection and design, in
different rock types that are suitable to different process routes. Strategic Mine Planning Conference, pp 133-136 (The Australasian
A trade-off study was developed in order to understand it better Institute of Mining and Metallurgy: Melbourne).
and tried to force the treatment of the whole orebody into just Randall, M, 1994. Application of advanced computer-based systems in
one of these process routes. Sometime this approach is valid, but open-pit evaluation, Mining Latin America, pp 77-87 (Institution of
usually it will be necessary to leave behind a significant portion Mining and Metallurgy: London).
of the reserve. It is also very usual to consider only one of the ore Stermole, F J and Stermole, J M, 2006. Economic Evaluation and
types and dismiss the other as waste or marginal ore. This is what Investment Decision Methods, 768 p (Investment Evaluations
normally happens on stand alone operations. This approach is Corporation: Colorado).
also valid, but it is not guaranteed that the reserve will be the

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 127
 HOME

Technogenic Iron Ore Formations in Siberia as a Secondary

Resource for Production of Ferrous, Rare, Non-Ferrous and
Noble Metals
P A Filippov1, S A Neverov2 and A A Neverov2

ABSTRACT In Kuzbass, the Sheregesh, Tashtagol, Kaz, Abakan

(underground mining), Teisk and Irbinsk (open cut mining)
Five mines and two concentrators supply raw materials to the
metallurgical complex in Siberia, whereas tailings of the iron ore mines are the main suppliers of iron ore to large West Siberian
concentrators are stockpiled at tailing dumps at the mine site. In Siberia, and Novokuznetsk metallurgical complexes. The run-of-mine ore
waste dumps reserve three billion tonnes of wastes and tailings is processed at a crushing and preparation plant to produce a
discharged by mines and ore preparation plants – their volume tends to primary iron ore concentrate of 38 - 44 per cent Fe in grade, then
increase by more than 15 Mt of overburden rocks annually. More than it is fed to the wet magnetic separation circuit with the yield of
3 Mt of the dry and wet magnetic separation tailings are discharged and the marketable product, viz the secondary iron ore concentrate of
stockpiled by mills and ore preparation plants. Open pit dumps keep 62 - 66 per cent Fe in grade at Mundybash and Abagursk ore
600 Mt of overburden rocks. Four underground mines accumulated 50 Mt preparation plants. Tailings are stockpiled at dumps at the mine
of wastes of 0 + 60 mm in size at tailing dumps. Wet magnetic separators site. In total, 373.8 Mt of raw iron ore were produced at Siberian
discharged more than 120 Mt of -2.0 + 0.045 mm slimes. Waste dumps
mines with the yield of 235.8 Mt of primary concentrate.
occupy more than 100 000 m2. Technogenic waste formations contain a
variety of valuable elements and compounds of commercial value. Production of the iron ore concentrates involves the formation of
Utilisation of secondary, technogenic raw materials is rather profitable, as large tailing dumps of 200 Mt, including tailings of preparation
the economic factors of the process are comparable with those for plants and slimes of the wet magnetic separation circuit (Table 1).
processing of mined mineral ores. Qualitative proportions of the valuable In Siberia, six iron ore mines built up 78.7 Mt of wastes from
content in dump wastes were evaluated based on geological exploration crushing and preparation plants in an area of 109 200 m2. The
of aged tailing samples by analytical and geophysical methods. By the large slime dumps are accumulated at ore preparation plants,
assay data, the iron ore tailings contain iron, cobalt, zinc, and other upgrading the primary mine concentrates into concentrates for
valuable metals, and the concentrator slimes have inclusions of precious smelters by employing the wet dressing and dust suppression
metal – gold and silver. The potential of technogenic iron ore reserves at processes. The dumps of Mundybash and Abagursk preparation
mines and ore preparation plants is substantial as a secondary source for
plants amount to 119.2 Mt of the wet separation slimes – 65 per
production of valuable metals in Siberia.
cent of the total wastes are -2.0 + 0.045 mm in size. At present,
the dump materials are classified and sold as construction gravel
INTRODUCTION or utilised as a fill to stow mining cavities under an agreement
Currently in Russia, more than 60 billion tonnes of industrial with the government supervising bodies.
wastes are stockpiled in surface dumps built up by operating
mines, smelters, heat power stations and other works. Most dump TABLE 1
materials are wastes from the mining industry (up to 90 per cent), Waste dumps at iron ore mines and ore preparation plants
smelter slags (two to three per cent), ashes of coal boilerhouses in Siberia.
(one to two per cent) and others (five to seven per cent). More than
75 per cent of mining wastes fall on overburden rocks at open pits Process Name Volume Area Valuable
(Mt) (m2) elements
and up to 15 per cent are tailings, discharged at ore, coal, or rock
preparation plants. Approximately three billion tonnes of rocks Mines
produced by opencast mining are processed annually, with more Underground Sheregesh 23.32 50 900 Fe, Mn, Zn
than 500 Mt of overburden rocks being mined at iron ore open mining, dry
Kaz 7.2 6300 Cu, Zn
pits. In total, about 800 Mt of tailings are stockpiled at dumps of magnetic
iron ore mines in the Siberian region. Mine dumps are increased separation Tashtagol 0.3 200 Fe
by more than five million cubic metres of overburden rocks Abakan 20.1 28 000 Co, Cu, Ni
annually – the crushing and concentrating factories discharge Open cut mining, Teisk 21.2 19 100 Co, Zn
above 2 Mt of tailings. This intensive mining activity results in a dry magnetic
heavy ecological impact on the local environment. More than 40 Irbinsk 6.7 4700 Co, Cu, Zn
separation
per cent of disturbed lands, 30 per cent of harmful atmospheric Ore preparation plants
emissions and up to 20 per cent of sewage water release into
natural water basins in Russia are because of mining operations. Wet magnetic Mundybash 30.1 60 100 Au, Ag, Zn
Of special note is the ecological load on the Kuznetsk Basin, one separation
Abagursk 89.1 385 200 ZN, Cr, Co
of the leading metal, coal, ferrous and non-ferrous producers in Total: 198.1 554 500 -
the Russian Federation.

The Institute of Mining of the Siberian Branch of the Russian

1. Senior Researcher, Institute of Mining, SB RAS, 54 Krasny
Prospect, 630091 Novosibirsk, Russia. Email: filippovpeter@mail.ru
2. Researcher, Institute of Mining, SB RAS, 54 Krasny Prospect,
630091 Novosibirsk, Russia. Email: neverovaa_79@mail.ru
3. Technogenic formation is a dump formed by many years of dumping
industrial wastes that could potentially have a commercial value to
produce new products by newly developed methods.
Academy of Sciences performed the survey studies aimed at
assessing the composition and properties of mine dump materials.
It was concluded by the exploration data that most dump materials
can be reprocessed and utilised as a secondary raw material
(Filippov, Uskov and Freidin, 2003; Gaidin et al, 2003, Freidin,
Filippov and Gaidin, 2002). Moreover, utilisation of technogenic3
raw materials can be economically profitable and comparable with
the processing of raw ore mined at geogenic mineral deposits

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 129
P A FILIPPOV, S A NEVEROV and A A NEVEROV

(Doroguntseva, 2006). The commercial evaluation of technogenic and ore preparation plant on a 15 - 20° slope, contains 1.6 Mt of
formations comprises a large-scale comprehensive exploration that ore wastes of variable iron content (Figure 1). The dump
includes both laboratory and in situ tests, as well as the stretches for 300 m, the maximum width of its horizontal section
development of standards for raw mineral materials, with reaches 40 m and the tailing bed is within 25 - 30 m. The iron
consideration for the summarised economic efficiency of the content in the Feofanovsky dump material is summarised in
utilisation of technogenic wastes. The ecological factor, involving Table 2. It is explicit that iron content in the upper layer of
preventing the release and discharge of hazardous components and consolidated tailings is higher by 12 per cent as compared to an
compounds into water basins and the atmosphere, should be average value, and tends to increase with the decrease in fraction
assessed along with the economic value of technogenic mineral size to 21.4 per cent at average iron content of 15.2 per cent. The
resources. Of particular economic and social importance is the fine fraction (zero to 15 mm) makes up one third of the total
restoration of disturbed lands in agricultural and recreation areas.
The magnetic properties of the aged iron ore tailings become
substantially weaker. The relatively high content of sulfide TABLE 2
minerals contributes to intensive oxidation processes followed by Iron distribution in fraction sizes.
liberation of gold, cobalt, nickel and other metals from the crystal Size fraction (mm) Ratio (%) Fe content (%)
lattice. These oxidation processes entail formation of ecologically
hazardous compounds of sulfur, arsenic and other elements. In Larger than 50 22.4 13.3
view of the above, processing of technogenic dump resources is 40 - 50 8.1 12.4
actualised through economic, ecological and social aspects. 20 - 40 20.5 13.2
The iron ore dumps at the Sheregesh Mine are of special
15 - 20 17.6 15.0
commercial value. Their reserves are above 20 Mt of tailings
with more than 20 per cent Fe in grade (Filippov, 2008). The 10 - 15 11.3 16.8
scientific and in situ investigations into production of the iron 5 - 10 16.3 19.2
ore concentrate from tailings of the Sheregesh Mine were
Less than five 3.7 21.4
performed at the Feofanovsky dump, favourably located in the
production mine site. The waste dump, 3 km from the crushing Total 100.0 Average 15.2

Fe content, %

FIG 1 - Isolines of iron content in mill tailings, Feofanovsky waste dump.

130 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 TECHNOGENIC IRON ORE FORMATIONS IN SIBERIA AS A SECONDARY RESOURCE
volume with iron content Fe = 19.1 per cent. The large size
fraction (>50 mm) constitutes one fifth of the total volume and is
relatively poor in iron – large-size lumps should be separated,
crushed and processed by dry magnetic separation.
CASE STUDY
The newly developed process flow sheet (Figure 2) was employed
in the full-scale tests for Feofanovsky-aged dump material at the
Sheregesh crushing-preparation plant. The tests were scheduled on
Sunday, when the factory exclusively processed dump material
delivered by trucks. The dump material was unloaded first into a
special bin at the factory, then fed into bunker No 36 in the second
technological line. The material was fed to separator Number 52
and then with no processing at the separator, it was transported by
conveyor lines Number 60, 49, 51, 53 and 55 to crusher Number
15 in the first technological line. The material ground down to 16
to zero millimetres was fed to screen Number 19, the oversize
fraction was re-crushed in fine crusher Number 23 and the
undersize was moved to conveyor Number 25. The re-crushed and
undersized products were combined on conveyor Number 25 and
fed to screen Number 27, which is equipped with three levels of
different-mesh sieves (20 × 20 mm upper, 18 × 18 mm middle and
13 × 13 mm lower sieves). After screening the oversize product of
+16 mm, it was conveyed by Number 51, 53 and 55 to crusher
Number 15, while the undersized product was conveyed by
Number 29 to magnetic separators Number 45 and 47. The
prepared feed was subjected to the dry magnetic separation at
separators at controlled operating modes. The final product
(primary concentrate) was conveyed by Number 61, 62 and 66 to
the finished product storage area. The tailings were conveyed by
Number 57, 59, 63, 65, 67, 71, 73 and 75 to screens Numbered 65
and 69 for classification to yield the finished construction material.
Thus, the primary iron ore concentrate of the required grade was
produced in the full-scale tests of four bulk samples of
Feofanovsky-aged dump material with the use of magnetic
separators Number 45 and 47 (Table 3).
By the test evidence, the intermediate product, viz a primary
iron ore concentrate of up to 36.9 per cent Fetotal in grade, was
produced under industrial conditions from old consolidated
tailings at the average concentrate yield of 20.8 per cent and
recovery of 48.3 per cent for separator Number 45. Concentrates
of 36.8 per cent Fetotal in grade with the average concentrate yield
of 26.8 per cent and recovery of 61.4 per cent were produced at
separator Number 47. The evaluated cost of one tonne of 33.4
divided by 36.9 per cent Fe concentrate produced from current
tailings of the preparation plant was US$3.2. The actual cost of
one tonne of the primary concentrate of 41 per cent Fe in grade
produced from a crude ore (26.4 per cent Fe) of the Sheregesh
underground mine was US$19.7. It is important to point out that
the cost of one tonne of 36.8 per cent Fe primary concentrate
produced from the ore of Teisky open pit is US$16.1. In Siberia,
the production capacity of crushing – preparation plants is
utilised by 80 - 95 per cent at all operating mines. Such a
TABLE 3
Assay data on the concentrate and tailings samples at Sheregesh preparation plant.
Separator Test Feed
number number
Fetotal (%) Femag (%) Fetotal (%)
1 15.9 8.9 36.4
Number 45 2 15.8 9.2 36.9
Average 15.8 9.1 36.7
1 15.3 9.3 36.8
Number 47 2 17.3 10.7 33.4
Average 16.3 10.0 35.1
Advances in Orebody Modelling and Strategic Mine Planning I
workload level no doubt hampers the processing of aged tailings.
The problem can be solved by employing a mobile modular
complex, designed to produce an iron ore concentrate from old
consolidated tailings (Figure 3). The mobile complex can be
mounted at the waste dump site. It consists of a receiving bin,
two one-drum magnetic separators, screening circuit for material
classification into size fractions and a belt-conveyor line for
transportation of concentrated products. The primary material
is unloaded by excavator, bulldozer or scraper facilities into a
receiving bin. The choice of loaders and trucks depends on the
capacity of a mobile concentrating complex and its operation mode.
The construction of the modular concentrating complex is
launched at the Sheregesh Mine to process 12 divided by 14 per
cent Fe tailings of Kumnagash dump, whose reserves are 9.1 Mt.
It is planned to produce an intermediate product (Fe = 35 per cent)
for further treatment at Sheredesh and Mundybash ore preparation
plants, with the yield of the finished concentrate (Fe = 42 per cent
or Fe = 62 per cent). The investment cost of the project is
US$820 000 including construction, assembly works and purchase
of mineral dressing equipment. The waste- concentrating complex
consisting of a concentrator module, diesel loader, five trucks and
a bulldozer, is served by a 25-man team. Its project production
capacity is 250 000 t of intermediate product and 130 000 t of the
primary concentrate at the annual planned processing rate of one
billion tonnes of wastes at Sheregesh preparation plant. The total
design production cost of the concentrate is US$8.7 per one tonne,
including US$3.2 per one tonne of the intermediate product at an
ore preparation plant, and US$4.4 per one tonne of the initial feed
at the concentrating complex. The project payback period is one
year for Kumnagash tailings. The process also involves the
production of a marketable by-product, namely, classified
construction gravel. Furthermore, the process provides higher
efficiency, more complete recovery of valuable components and
less ecological impact, thanks to the optimal operating modes of
magnetic separation of iron ore tailings, heterogeneous in
granulometric composition and content of valuable components.
The exploration survey of Siberian mines revealed the adequacy in
their qualitative characteristics and proved the satisfactory
reliability in interpretation of the Fe recovery data obtained for
one waste dump to others.
The Mundybash slime dump accumulates 30 Mt of iron
bearing slimes of wet concentration and the fine-grain material
sized -2.0 + 0.045 mm makes up 65 per cent of the total mass. In
evidence derived from the comprehensive analysis, Zn, Cu, Co
and Mn elements prevail in the slimes. Of great value is the noble
metal (gold, silver) content in slimes – these metals are distributed
irregularly around the dump area and their content in core
specimens varies – Au = 0.1 - 1.2 g/t (weighted average value is
0.616) and Ag = 0.5 - 7.0 g/t. The centrifugal separation provides
the maximum gold recovery of 93 per cent at the gold
concentration ratio of 8:23 versus the initial gold content in a
sample under laboratory conditions. The process is applicable to
recover silver, sulfide concentrates and finely-dispersed
construction materials along with the base gold.
Concentrate Tailings
Femag (%) Yield (%) Recovery (%) Fetotal (%) Femag (%)
30.3 22.9 52.4 9.8 3.9
30.0 19.2 44.8 10.8 2.9
30.2 20.8 48.3 10.3 3.4
31.3 22.7 54.6 9.0 1.8
27.8 31.8 61.4 9.8 2.3
29.6 26.8 57.7 9.4 2.1
Spectrum Series Volume 17 131
P A FILIPPOV, S A NEVEROV and A A NEVEROV

FIG 2 - The in situ test flow sheet for processing of current tailings of Sheregesh crushing-preparation plant.

132 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 TECHNOGENIC IRON ORE FORMATIONS IN SIBERIA AS A SECONDARY RESOURCE

Wastes CONCLUSIONS
1 In Siberia, about 200 Mt of iron ore dressing process wastes are
stockpiled – the wastes currently occupy an area of more than
100 000 square metres. It has been found that technogenic
2 formations of aged mine wastes and tailings contain potentially
valuable ferrous, non-ferrous and noble metals (Fe, Co, Zn, Au
and Ag). The reasonability of development and implementation
4 of innovative projects for the processing of technogenic waste is
3 apparent, as modern mineral-dressing processes allow the
efficient processing of available dump materials to produce
marketable products: iron ore concentrate, base gold, sulfide
concentrates, and construction materials.
The planned concentrator for recovery of gold and sulfide
5 concentrates from current and aged tailings is based on
the exploration studies carried out at the Mundybash ore
preparation plant.

6 REFERENCES
Doroguntseva, N P, 2006. Development of comprehensive utilisation of
natural and secondary mineral resources, based on diversification of
activity of mining enterprises (in Russian), PhD thesis, Novosibirsk.
Filippov, P A, 2008. Potentials of technogenic formations in mines of
5 West Siberia, Journal of Mining Science, 4:386-390.
2 Filippov, P A, Uskov, V A and Freidin, A M, 2003. Perspectives for
processing of technogenic raw materials at the Sheregesh Mine (in
Russian), in Science Intensive Processes for Mining and Beneficiation
of Mineral Resources, pp 164-166 (IGD SO RAN: Novosibirsk).
Freidin, A M, Filippov, P A and Gaidin, A P, 2002. Process for treatment
of iron ore tailings, Russian Federation Patent No 2190027.
Construction material Concentrate (Fe=35%) published in Bulletin of Inventions, No 27.
Gaidin, A P, Filippov, P A, Komissarov, A V and Onofreichuk, V Ya,
2003. Treatment of old consolidated tailing dump materials in the dry
1 - Receiving bin; 2 - Doser ; 3 - Screen ; magnetic separation cycle at a concentrating factory of OAO
4 - Magnetic separator; 5 - Conveyor ; 6 - Discharge bin . ‘Sheregesh Mining Board’ as a factor improving the regional
ecological situation and economic efficiency of iron ore processing,
in Proceedings Fourth Congress of Mineral Processing Experts of
FIG 3 - Modular complex for processing of tailings. CIS Countries, pp 192-193 (MISiS: Moscow).

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 133

Advancing Strategic Mine Planning – A Software Perspective
C Tonkin1
ABSTRACT
Strategic mine planning for open pit mines is a mature process, in which
sophisticated software tools are used to optimise everything, from pit
design, to pushbacks, to scheduling. Since inputs into the planning
process are complex and uncertain, current practices call for breaking the
task into component parts and addressing them sequentially. Research
suggests that this paradigm may soon change, with the expanded
application of such concepts as simultaneous planning. In addition,
techniques for managing the unknown inputs of metal content and future
financial conditions – notably stochastic simulation and real options –
promise to leverage uncertainty in order to deliver increased mine value.
While to date techniques focus primarily on open pit mines, they lend
themselves to a logical extension into the underground mining environment
as well.
INTRODUCTION
Strategic mine planning is the art of designing a business strategy
that will deliver maximum business value in extracting an
available resource, through the optimal allocation of time,
capital, technologies and people. For more than two decades, the
global mining industry has used increasingly sophisticated
mathematical methods and techniques implemented in software
packages to optimise this planning process. However, strategic
mine planning depends on a complex, interdependent, and highly
uncertain system of inputs. Some uncertainties relate to the
resource itself, such as grade and tonnage of material or metal in
the ground, and geotechnical stability of the resulting excavation,
and others including uncertainty in the operating conditions that
will prevail during production and capital required. We also live
with the fact that, although a significant upfront investment is
required to start a mining project, the resource itself will be
extracted over a period of five, ten, 20, or 30 years, or even
longer. The return on investment will therefore be delayed into
future and uncertain times.
A key challenge presently facing the mining industry lies in
optimising the mine as an integrated process. Ideally, all relevant
factors should be considered and addressed simultaneously, from
the perspective of understanding how they influence the strategic
objective for the business – typically maximum net present value
(NPV) for the mine. What type and grade of material should be
extracted? When is the best time to extract this material? Where
is the best place to begin mining? How should the mine be
staged, such that the strategy can be readily modified if
conditions change? Can the various elements in the deposit be
blended to create an optimal product mix for the market? What
risk reduction strategies should be followed – mining a smaller
extraction shape first to minimise the impact of future price
uncertainty, carrying hedge books, improving data inputs through
increased drilling density? These questions and more relate to
optimising the mining process, and a similar number of questions
can be raised for the optimisation of the downstream processing
and logistics for the mine product.
Because future business and market conditions can only be
estimated, planning must be based on the best possible
predictions related to cost, price, and other key factors. Today’s
strategic mine planning practices do a good job, with excellent
contributions on both the technological and business side.
Current software tools make the planning process faster, easier,
1. Technical Product Manager – Scheduling and Optimisation,
Gemcom Software International Inc.
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
and better able to handle increasingly complex problems. We
have come a long way along the optimisation pathway – for
example, expanding pit optimisation to cut-off and blend
optimisation – and we have a good understanding of the business
processes related to strategic mine planning.
This paper provides the context surrounding strategic mine
planning and analyses current techniques that are applied to the
problem space. Detailing the history of advances in the area to
date, the paper describes innovative solutions and their
relationship to the complexity of the mine planning problem.
Finally, a number of technologies are discussed for advancing
strategic mine planning in the future.
The complexity of strategic mine planning
Strategic mine planning is concerned with the overall direction
of the mining company. Based on a clear understanding of the
market and relevant economic factors (price, cost, and discount
rate assumptions), it seeks to optimise the allocation of scarce
capital resources by aligning physical parameters such as the
grade and tonnes or metal produced with the requirements of the
business. Business requirements might include positive cash
flows throughout the mine life, a hurdle rate for investment that
must be satisfied, a minimum amount of metal to be delivered at
a given time, or a minimum grade/quality target for the final
product.
Some key questions related to the mining operation include:
• What capital configuration and associated cost model will
maximise the return on investment?
• What are the implications of different financial and slope
scenarios, and should the final design be adjusted to suit?
• To what parameters is the deposit most sensitive, and is
further investment needed before commencing the mining
operation?
Strategic mine planning addresses these questions and more,
enabling the mine planner to make decisions about the final
strategy and ultimate mine configuration based on sound
analysis.
Pit optimisation alone does not provide sufficient information
on the mine plan to enable sound decision-making. The final
mine strategy decision goes well beyond the size and shape of
the final pit by looking at the mine as a whole – pit, schedule,
sequence, staging, plant, port, ore waste discrimination,
blending, and pushback selection. By identifying a range of
viable options, investment decisions can be made according to
the requirements of the individual situation. To do this
effectively, a range of advanced planning tools and rapid
evaluation of different scenarios are required.
Many different planning options are available, depending on
the size of the final pit, as shown by Whittle (1997). Figure 1
depicts several possible strategies, including maximum resource
point, maximum NPV point, and a range of less risky designs
requiring a smaller capital outlay. The shaded area in the diagram
represents all financially viable and technically feasible mine
designs. Looking at the far right side of the graph, the decision to
extract the maximum economically viable reserves represents an
aggressive approach. The risk-neutral mine design at the top of
the curve would exploit the resource for the purpose of returning
maximum NPV to investors, sacrificing total ore tonnes. Farther
left is a range of risk-averse designs that reduce capital investment
Spectrum Series Volume 17 135
C TONKIN
FIG 1 - Making resource allocation decisions.
and mine life but still return a positive NPV. It is important to
note that any of these alternatives might be considered optimal,
depending on a company’s resource allocation strategy. In this
example, the information needed to make these critical resource
allocation decisions, along with drill-down capabilities that
provide more detail on the capital and operating cost, price, and
recovery inputs for any given option, was generated in Gemcom
Whittle™ software. Having this level of detail visible, accessible,
and auditable allows the mine planner to rapidly analyse the
different options and make the best choice for a given case.
The extraction strategy that is selected will have a direct impact
on the cash flow of the mine. In Figure 2, for example, the
decision has been made to maximise NPV within the constraints
of the final pit design. The graph shows the effect of different
sequencing approaches, with the two lines representing two
different extraction strategies with exactly the same final pit,
resources, ore, and waste tonnes. Depending on the extraction
strategy, the project will deliver vastly different values. Strategic
mine planning is a complex process, and the many variables
preclude a simple mathematical solution. Mine life, NPV, return
FIG 2 - Identifying the best strategy for any given decision.
136 Spectrum Series Volume 17
on investment, reserve targets, grade consistency, geotechnical
uncertainty, capital investment, commodity prices, stockpiles –
all these and more make planning optimisation an ongoing
challenge.
Current strategic mine planning practices lay the
foundation for their advancement
The standard approach to strategic mine planning is to break the
overall task into discrete, manageable pieces. For example, mine
planners look at different slope scenarios, using arc relationships
between blocks in the model, use costs and revenues to
determine ore waste discrimination and block values, and employ
Lerchs Grossman (Lerchs and Grossman, 1965; Whittle, 1988)
pit optimisation routines to create a range of options for mining
the orebody. Next comes strategic scheduling where planners
typically follow a step-wise process of pushback selection,
staging the mine to expose the ore, balance stripping requirements,
and improve return on investment. They seek to maximise NPV
at the marginal cut-off, then investigate further value by looking
at cut-off optimisation and blending strategies.
The development timeline of Gemcom Whittle, a long-time
leader in strategic mine planning software, provides insight into
the industry trends that have led to the practices and techniques
that are widely used today. Whittle brought the first commercial
three-dimensional pit optimisation package to market in 1985 –
20 years after publication of the Lerchs Grossman pit
optimisation theory (Lerchs and Grossman, 1965) – allowing its
users to quickly and accurately calculate the optimal pit for a
given resource. This hitherto unavailable ability to calculate the
optimal pit through an automated process provided mine
planners with both a mathematically optimal solution and an
enormous time saving compared to manual processes. Using
Whittle for pit optimisation, the mine planner now had a process
that provided the maximum value solution for the final pit. This
enabled more time to be spent on accurately modelling costs,
developing alternative scenarios and, for the first time, analysing
the sensitivity of the deposit to changes in input parameters, most
importantly commodity prices. (Whittle, 1988)
Advances in Orebody Modelling and Strategic Mine Planning I
 ADVANCING STRATEGIC MINE PLANNING – A SOFTWARE PERSPECTIVE

The following year, Whittle added the fourth dimension, time,

which enabled discounted cash flow to be calculated. This
enhancement enabled the mine planner to make decisions that
would improve the NPV and return on investment for a deposit. In
1995, Whittle incorporated Lane’s cut-off optimisation theory
(Lane, 1988; Whittle and Wharton, 1995), according to which cut-
offs are varied over time to maximise the NPV for a mining
sequence. By ‘productising’ published theories through commer-
cially available software, Whittle thus created a core set of
building blocks for strategic mine planning – pit optimisation,
economic analysis, and cut-off optimisation – that delivered the
capability to determine optimal solutions for a deposit and to
examine different mining scenarios and sensitivities. The years on
either side of the millennium – the late 1990s and early 2000s –
marked a new stage in the industry, with innovative solutions join-
ing the basic building blocks to deliver the benefits of Whittle to a
wider range of projects and unlock even greater value with
additional optimisation routines. Whittle development continued
to reflect the needs of the industry, beginning with the introduction
of multi-element deposit capability. This feature clearly showed
revenue contributions from more than one element, giving poly-
metallic deposits the same benefits enjoyed by single-element
deposits, that is, pit optimisation, economic analysis, and cut-off
optimisation. To enable mine planners to achieve maximum NPV,
Whittle developers undertook a unique research initiative resulting
in a proprietary schedule optimisation technique through
its Milawa Algorithm module (Wharton, 2000). The Milawa
Algorithm, results output example shown in Figure 3, provides
automated schedule optimisation routines that maximise the NPV FIG 4 - Whittle’s Navigation Tree organises projects into an
for a given final pit. With all the complexities involved in strategic easy-to-use structure.
mine planning and the utilisation, manipulation, and management
of large amounts of data and end results, software ease of use, the enables users to generate the information they need and even test
logical organisation of Whittle projects is shown in Figure 4, hundreds of different optimisation scenarios quickly, in order to
became the next area addressed by Whittle. find the one that will deliver the best results. Today, Whittle is
Considerable research went into developing the unique, widely known for its ease of use. Following this, Whittle’s
logically organised project environment of the software, which Blending, Stockpiles and PushBack Chooser modules came to

FIG 3 - Milawa schedule to maximise net present value.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 137
C TONKIN
market in rapid succession. These modules enabled more
detailed optimisation and analysis of the mining process faced by
companies. The Blending module delivered the ability to
optimise blended final products or to blend the input to the
processing plant (Wharton, 2004). The PushBack Chooser
module automatically selected the optimal NPV set of pitshells
for any given final pit (Figure 5). The Stockpiles module
delivered the capability to model the use of stockpiles in various
ways, a critical part of strategic mine planning for some
operations. Now, the industry had automatic scheduling routines
to optimise the sequence for NPV, not just report it, and to
optimise for a range of commodities, final products, and
pushback and stockpile strategies. The software was then
expanded to work with data from different systems, further
improving the user experience. Working from the building blocks
in the previous stage, the millennium years saw the creation of a
total solution for open pit mine planning.
From 2003 to the present, Whittle and the mining industry have
stretched the limits of strategic mine planning to solve ever more
complex problems. A raft of features have added new value to the
market, including Whittle’s Multi-Mine module to optimise not
just a pit, but an entire mine operation consisting of multiple pits
and all related products and processes (Hall, 2004). Optimising
multiple mines simultaneously helped exploit synergies between
mines, making it possible to maximise NPV, balance stripping
requirements, and achieve grade targets. To tailor the software to
individual company needs, the Value Expressions module
FIG 5 - Create net present value practical pushbacks in a
single step.
FIG 6 - Timeline for Gemcom Whittle.
138 Spectrum Series Volume 17
expanded the methods available for calculating the value of a
block. Companies could now use predefined block values or,
alternatively, custom calculations to value blocks. This capability
provided virtually unlimited ways in which the software could be
applied, and the level of complexity it could model and optimise.
Value expressions unlocked the potential for pit optimisation to
work on dynamic block values.
Next came high-performance binary file support for growing
data sets, plus a series of new technologies designed to expand
memory and limits, streamline processes, support block
calculations, maximise NPV, and more. The drive to deliver
optimal solutions, combined with practical mining scenarios, has
culminated in developments such as the NPV Practical PushBacks
module. This module automatically creates a set of pushbacks
that address maximum NPV and minimum mining width
requirements, providing fast and accurate methods for optimising
NPV with more practical constraints. In summary, Whittle’s
advancements have responded to the industry’s demand for more
sophisticated tools that can be applied to an ever wider range of
strategic mine planning challenges and project types. The focus
has remained on providing mathematically optimal solutions to
meet the real-world challenges faced by mining companies.
Moving into the future
Looking at the Whittle example, it is clear that strategic mine
planning theory has been on a trajectory of virtually exponential
development, as shown in Figure 6.
This trajectory is good news for the mining industry, because it
means we should be able to look forward to significant advance-
ments in the near future. As we have seen, today’s strategic mine
planners generally break problems down into smaller pieces,
optimise discrete components, and solve individual parts. By
doing so, they are able to exploit mineral deposits quite effectively.
But as our ability to model more realistic scenarios grows, and as
our input models become increasingly complex and inter-
dependent, this sequential approach will fall short when it comes
to optimising mine value. What are some of the current research
and possible future technologies that may take the mining industry
to the next level? One exciting research area looks at simultaneous
planning of building blocks, making it possible to take a more
holistic approach to the overall planning process. The goal is to
integrate each building block and optimisation process to solve the
mine planning process simultaneously.
Research in this area continues, using frameworks such as new
Mixed Integer Linear Programming (MILP) models (Ramazan,
2007; Stone et al, 2007) and global optimisation techniques
(Whittle, 2009). Limited commercial solutions for simultaneous
planning already exist: Some Gemcom clients in the coal
industry have been using a commercial MILP solution for a
Advances in Orebody Modelling and Strategic Mine Planning I
 ADVANCING STRATEGIC MINE PLANNING – A SOFTWARE PERSPECTIVE
FIG 7 - Simultaneous planning.
number of years. While current technologies go some way
toward solving mine planning processes simultaneously, they all
take pit optimisation as an input to the process and involve other
building blocks to varying degrees. Moving forward, solutions
will need to deal with schedules, stockpiles, ore waste
discrimination, and blending with even greater power and
sophistication. Simultaneous planning through technologies such
as MILP solutions and global optimisation are important
developments, which the industry can hopefully look forward to
exploiting in the coming years.
Leveraging uncertainty
Other research focuses on the uncertainty and risk inherent in
strategic mine planning, for example in geological, geotechnical,
operational, and price models. Uncertainty and risk can be
reduced by improving the inputs to the models – more drilling
would provide a better assessment in the geological model, for
example - while additional drilling and analysis would enhance
the accuracy of geotechnical predictions. Reducing uncertainty
in this way is costly and reaches the point of diminishing return;
at best, it can reduce but not eliminate the risk.
Still further research suggests the use of techniques, built on
foundations of sensitivity and scenario analysis, for working with
uncertainty. These types of techniques can help to evaluate the
consequences of uncertain inputs. Grade and metal uncertainty
may also be evaluated using conditionally simulated models,
quantifying the risk associated with a given mine plan or
choosing mine plans that operate well under uncertainty; hybrid
pit techniques and risk-based optimisation frameworks are
examples of this approach (Whittle, 2007; Dimitrakopolous,
Martinez and Ramazan, 2007). Current research into simulated
annealing and stochastic integer programming methods for
production scheduling (Godoy and Dimitrakopolous, 2007;
Ramazan and Dimitrakopolous, 2007) also offers the ability to
leverage uncertainty in mine planning. Using uncertainty to drive
mine planning decisions represents a leap forward from creating
mine plans based solely on deterministic inputs.
What about uncertainty in operational and financial models?
Research in this area takes its cue from the financial community in
the form of the ‘real options’ evaluation technique, which
recognises that managerial decisions can and will be taken in the
future to react to changing circumstances. This method essentially
quantifies the value of management flexibility in a world of
change and uncertainty. Mine plans that permit different options
may have higher value than those that do not; in other words, the
option to take certain actions has a value in and of itself. When we
Advances in Orebody Modelling and Strategic Mine Planning I
examine a solution space of many possible outcomes based on
multiple realisations of the orebody, for example, real options can
be used to help determine the best candidate. (Groeneveld, Topal
and Leenders, 2009; Abdel Sabour and Dimitrakopolous, 2010, in
this volume). While price is the biggest influence on profit, it is
not constant. How can we model future prices and discount rates?
Current practice reduces the price into the future, in order to
mitigate some of this uncertainty. An alternative is to use option
techniques to model the effect of changing the mine plan in
response to future, uncertain price movements.
Moving underground
Strategic mine planning has focused primarily on open pit
environments so far, but there is a logical extension to
underground mining as well. In Whittle, for example, there are
already techniques to help determine the transition point between
surface and underground mining. Providing information to the pit
optimiser that an underground opportunity exists, influences the
decision-making regarding final pit size. Moving to the
underground mine itself, the opportunity exists for more planning
tools to optimise various parts of the mine planning process. The
strategic mine planning process for underground is similar to that
for open pit. It considers the size and shape of the final resource,
the ore waste discrimination point, the best way to stage the mine
and extract the resource, and finally the optimisation of the
mining sequence itself. As with open pit, individual parts of the
underground process – for example, stope sizing and decline
placement – can be optimised separately. Currently, this planning
is typically done manually, although Gemcom PCBC™ provides
optimisers for block caving operations, including cave footprint
finders and a schedule optimiser. Current research in the industry
is looking at ways to optimise each step in the mine planning
process (Brazil et al, 2009), filling in the gaps with manual
techniques where the tools do not yet exist. This is similar to the
building blocks stage of open pit strategic mine planning.
Research projects are now examining ways to apply some open
pit mine planning advances to underground mining: MILP and
other formulations can be used to optimise schedules for
underground (Elkington, Durham and Myers, 2009) and metal
uncertainty techniques also have potential applicability (Grieco
and Dimitrakopolous, 2007). Since underground mining can
leverage the advancements made for open pit, it seems reasonable
to expect that the timeline should be significantly condensed.
CONCLUSION
It has taken 30 years to evolve from manual plans to our current
level of sophistication in open pit strategic mine planning. Many
excellent tools and techniques are now available, but they tend to
be sequential and siloed. In the near future, it is likely that the
industry will start leveraging sophisticated methods to incorporate
uncertainty into decision-making processes, along with simul-
taneous planning technologies. Underground strategic mine
planning, while currently not as advanced as open pit, is expected
to move along the same path at a more advanced rate.
Regardless of future direction, however, one basic tenet will
not change: Within the strategic mine planning framework,
flexibility must be ensured to determine different mine plans
based on business objectives, as not every company will opt for
maximum NPV. Robust mine plans, managerial flexibility, and
innovative software solutions make this a truly exciting time to
be in the mining industry.
REFERENCES
Abdel Sabour, S A and Dimitrakopolous, R, 2010. Accounting for joint ore
supply, metal price and exchange rate uncertainties in mine design, in
Advances in Orebody Modelling and Strategic Mine Planning I (ed: R
Dimitrakopolous), pp 29-34 (The Australasian Institute of Mining and
Metallurgy: Melbourne).
Spectrum Series Volume 17 139
C TONKIN

Brazil, M, Grossman, P A, Lee, D H, Rubenstein, J H, Thomas, D A and Ramazan, S, 2007. Large-scale production scheduling with the
Wormald, N C, 2009. Access optimisation tools in underground mine fundamental tree algorithm, in Orebody Modelling and Strategic
design, in Proceedings Orebody Modelling and Strategic Mine Mine Planning, second edition (ed: R Dimitrakopolous), pp 121-127
Planning (ed: R Dimitrakopolous), pp 237-241 (The Australasian (The Australasian Institute of Mining and Metallurgy: Melbourne).
Institute of Mining and Metallurgy: Melbourne).
Ramazan, S and Dimitrakopolous, R, 2007. Stochastic optimisation of
Dimitrakopolous, R, Martinez, L and Ramazan, S, 2007. Optimising open long-term production scheduling for open pit mines with a new
pit design with simulated orebodies and Whittle Four-X – A integer programming function, in Orebody Modelling and Strategic
maximum upside/minimum downside approach, in Orebody Mine Planning, second edition (ed: R Dimitrakopolous), pp 385-391
Modelling and Strategic Mine Planning, second edition (ed: R (The Australasian Institute of Mining and Metallurgy: Melbourne).
Dimitrakopolous), pp 201-206 (The Australasian Institute of Mining
Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse,
and Metallurgy: Melbourne).
P, 2007. Blasor – Blended iron ore mine planning optimisation at
Elkington, T, Durham, R and Myers, P, 2009. Optimising value for an Yandi, Western Australia, in Orebody Modelling and Strategic Mine
underground project configuration, in Proceedings Orebody Planning, second edition (ed: R Dimitrakopolous), pp 133-136 (The
Modelling and Strategic Mine Planning (ed: R Dimitrakopolous), Australasian Institute of Mining and Metallurgy: Melbourne).
pp 249-255 (The Australasian Institute of Mining and Metallurgy:
Wharton, C, 2000. Add value to your mine through improved long term
Melbourne).
scheduling, in Proceedings Whittle North American Strategic Mine
Godoy, M and Dimitrakopolous, R, 2007. A multi-stage approach to Planning Conference, Breckenridge.
profitable risk management for strategic planning in open pit mines,
Wharton, C, 2004. The use of extractive blending optimisation for
in Orebody Modelling and Strategic Mine Planning, second edition
improved profitability, in Proceedings Orebody Modelling and
(ed: R Dimitrakopolous), pp 337-344 (The Australasian Institute of
Strategic Mine Planning (ed: R Dimitrakopolous), pp 69-76 (The
Mining and Metallurgy: Melbourne).
Australasian Institute of Mining and Metallurgy: Melbourne).
Grieco, N and Dimitrakopolous, R, 2007. Grade uncertainty in stope
Whittle, D, 1997. The relationship between economic design objectives
design – Improving the optimisation process, in Orebody
and reserve estimates, in Proceedings Resource to Reserve Inputs
Modelling and Strategic Mine Planning, second edition (ed: R
Seminar, pp 27-29 (The Australasian Institute of Mining and
Dimitrakopolous), pp 167-174 (The Australasian Institute of Mining
Metallurgy: Melbourne).
and Metallurgy: Melbourne).
Whittle, D and Bozorgebrahimi, A, 2007. Hybrid pits – Linking
Groeneveld, B, Topal, E and Leenders, B, 2009. A new methodology for
conditional simulation and Lerchs-Grossman through set theory, in
flexible mine design, in Proceedings Orebody Modelling and
Orebody Modelling and Strategic Mine Planning, second edition (ed:
Strategic Mine Planning (ed: R Dimitrakopolous), pp 109-117 (The
R Dimitrakopolous), pp 323-328 (The Australasian Institute of
Australasian Institute of Mining and Metallurgy: Melbourne).
Mining and Metallurgy: Melbourne).
Hall, G, 2004. Multi-mine better than multiple mines, in Proceedings
Whittle, J, 2009. The global optimiser works – What next? in
Orebody Modelling and Strategic Mine Planning (ed: R
Proceedings Orebody Modelling and Strategic Mine Planning (ed: R
Dimitrakopolous), pp 25-30 (The Australasian Institute of Mining
Dimitrakopolous), pp 3-5 (The Australasian Institute of Mining and
and Metallurgy: Melbourne).
Metallurgy: Melbourne).
Lane, K F, 1988. The Economic Definition of Ore: Cut-off Grades in
Whittle, J and Wharton, C, 1995. Optimizing cut-off grades, in
Theory and Practice, 149 p (Mining Journal Books: London).
Proceedings 1995 Optimizing with Whittle: Strategic Mine Planning
Lerchs, H and Grossman, I F, 1965. Optimum design of open pit mines, (Whittle Programming Pty Ltd: Melbourne).
CIM Bulletin, volume 68 (Canadian Institute of Mining and
Metallurgy).

140 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I

Integrated Strategic Planning at EKATI Diamond Mine
B Coutts1
ABSTRACT
Effective strategic planning at EKATI Diamond Mine, Northwest
Territories, Canada, requires both the engagement of the business into the
process and the integration of the results into deliverable schedules and
developments. The goal is to position EKATI as the lowest cost premier
producer of diamonds in the developed world. This will be accomplished
by maximising EKATI’s economic and environmental sustainability
through identifying and developing options with the potential to deliver a
robust, long life, high quality asset. To achieve this, strategic planning must
integrate directional policies with the physical input ranges applicable at
EKATI via an annual, comprehensive, inter-linked planning system to both
optimise and challenge the present state. Diamonds present unique
uncertainties to the planning process. The evaluation of variance in
certainty and range within the multiple inputs enables the assessment of
strategic opportunities and risks. The development and application of an
EKATI specific evaluation and optimisation model facilitates the rapid,
objective application and assessment of multiple strategic policies,
constraints and development options. Effective communication and
integration of strategic planning results into deliverable operational
schedules is critical to both maximise the value of EKATI and execute
opportunities to grow a long-term sustainable business.
INTRODUCTION
BHP Billiton Diamonds Inc is the operator of the EKATI
Diamond Mine, which is located approximately 300 km
northeast of Yellowknife and 200 km south of the Arctic Circle
in the Northwest Territories of Canada. EKATI’s environment is
sub-Arctic desert, where the average annual daytime high is -9°C
and permafrost occurs to depths of 330 m (Figure 1). The mine
operates as a fly-in, fly-out camp with 640 rooms and often >800
people on-site. The major route for annual re-supply and moving
heavy loads is the seasonal winter road, operated over frozen
lakes for approximately two months a year and shared with two
other diamond mines. Following the initial discovery of the
Lac de Gras diamondiferous kimberlites in 1991, over 150
kimberlites have been identified within the EKATI lease
holdings. Since opening in 1998, EKATI has produced in excess
of 40 Mct of diamonds (to end FY08) from a combination of
open pit and underground mining of six kimberlite pipes. This
paper is a case study of the current strategic planning processes
and systems at EKATI, the intended future developments to these
systems and the effective wider business integration of both
inputs and results.
The planning cycle applied at EKATI is a BHP Billiton
mandated procedure which allows for aligned planning
throughout the company. This procedure is mandated by Group,
delivered by the Customer Sector Group (CSG) and executed by
the Asset. In this case the CSG is Diamond and Specialty
Products (D&SP) and the Asset is EKATI. The planning cycle is
an annual integrated process that links high level directional
planning with detailed delivery planning (BHP Billiton, 2008). It
starts with the Chief Executive Officer’s (CEO) Message, which
reinforces the elements supporting the directional strategy. The
directional planning phase considers full potential, key strategic
choices and imperatives evaluated through the Resource
Development Potential (RDP) and Life of Asset (LOA) planning
procedures. The delivery planning phase focuses on resource
allocation improvement initiatives and action and includes the
1. Superintendent Long Range Planning and Projects, BHP Billiton
Diamonds Inc, 1102 - 4920 52nd Street, Yellowknife NT X1A-3T1,
Canada. Email: Ben.P.Coutts@bhpbilliton.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
5 Year Plan (5YP) and 2 Year Budget processes. This paper will
focus on the RDP which is the key strategic element of the
planning cycle. The RDP applies inclusive, divergent thinking at
the directional phase of the planning cycle to understand the
Asset’s full potential value. Key sources of value during the RDP
process are realised through the integration of strategic direction
with physical input ranges, an investigation of value and business
opportunities through unconstrained thinking and the provision
of a development plan to maximise full potential return at Asset
level, including the value of exercising growth options. This
process enables sound decisions through structured analysis of
the facts and facilitates the evaluation of uncertainty and its
resultant impacts upon growth options.
The effective communication and integration of strategic
planning results within the planning cycle is achieved through
structured output which includes clearly identified and
constrained go-forward frameworks, approved by CSG
management, which form the basis for further evaluation during
the planning cycle. Additional deliverables are a comprehensive
and prioritised list of development projects to move towards full
asset value and the identification of potential development and
improvement options including trigger values to move into their
execution.
Approval for the development of the EKATI mine was justified
on the basis of the 1996 Feasibility Study (BHP Diamonds Inc,
1996). Since that seminal assessment, regular re-evaluations of
the EKATI strategic plan have provided updated economic
parameters upon which future business decisions have
been made. Incorporated into these re-evaluations are changes
to multiple physical parameters, increases in operational
experience, a dynamic financial environment, paradigm shifts in
marketing and the constant evolution of modelling tools. As such
this paper presents a selective snapshot of a continuous process.
INPUTS
The key strategic imperative for EKATI is to unlock its long-term
potential in a manner that optimises the return to all stakeholders.
This can be accomplished by maximising EKATI’s economic
and environmental sustainability through identifying and
developing options with the potential to deliver a robust, long
life, high quality asset. To achieve this, the RDP must integrate
directional strategy with the physical input ranges applicable at
EKATI.
Directional strategy
A clear and thorough determination of the present state of
EKATI is a critical component in the preparation of the RDP.
Processes undertaken include performing a constraints analysis,
conducting a value driver analysis, reviewing EKATI’s risk
register and updating safety, environmental and regulatory
requirements. Constraints analysis is a structured approach to the
analysis of constraints or ‘bottlenecks’ on the output of the
operation. Value driver analysis is conducted in order to
determine the key variables which affect the value of the
operation. The risk register lists identified and understood risks
to the entity. The active engagement of EKATI’s management
team is fundamental to achieving consensus in these analyses.
Development of EKATI’s required forward looking strategic
framework is achieved through the collation of all available
strategic inputs including the CEO message, the CSG strategic
Spectrum Series Volume 17 143
B COUTTS
FIG 1 - EKATI location.
framework (Figure 2) and scenarios. Strategic scenarios are
plausible, widely divergent but internally consistent models of
possible future worlds. Each may include a set of future macro
economic conditions and an assessment of likely implications as
to the market, costs, competitors and environmental issues.
Review of these materials by EKATI’s management team
generates the scope of the RDP. Essential to the directional
preparation is the collection and review of current Research and
Development (R&D) initiatives that could have a future impact
upon EKATI. Current R&D initiatives under evaluation at
EKATI include potential alternate methodologies for mining,
energy and transport. Given the fact that EKATI’s product is sold
into a complex and opaque global market, the direct involvement
of marketing is vital to enable the assessment and review of
market trends, forces driving change, marketing innovations and
product concepts. As EKATI’s RDP must incorporate the full
potential mineral inventory, consultation with the exploration and
resource development group ensures the inclusion of all potential
greenfield and brownfield exploration results. Amalgamation of
the directional and physical elements is achieved through a
facilitated problem framing workshop. A problem framing
workshop commences with the current state, clearly and
succinctly states the problem to be resolved, applies structured
decision hierarchies to generate both strategic decisions within
potential settings and divergent consistent themes for further
evaluation. The workshop participants ideally include the
broadest possible range of contributors of information,
decision-makers and output owners. Participation of EKATI’s
management team is essential to ensure the resultant strategic
themes, decisions and alternatives provide an accurate and
comprehensive direction to the RDP.
144 Spectrum Series Volume 17
Physical input
Current strategic planning at EKATI optimises a mineral inventory
of 12 physically discrete and widely dispersed kimberlite
orebodies developed by open pit and/or underground operations.
Each kimberlite pipe exhibits unique value and recovery
characteristics in combination with specific capital and operating
requirements. Blended feed is processed in a single plant with
multiple potential operating profiles. All the physical inputs
evaluated in the RDP are subsequently applied in further
progressive stages in the planning cycle. Ensuring consistency of
value in these inputs is fundamental to achieving accurate and
comparable results from the planning cycle. Achieving this
requires constant vigilance and communication within and
between the many technical teams involved. Currently EKATI
seeks to regulate this through controlled sign-off sheets of required
values. Development of an integrated planning database intended
to further facilitate both the accurate generation and secure
provision of input values is discussed further in the progression
section.
Resources
As discussed in Dyck et al (2004) and Kleingeld and Nicholas
(2007) the modelling of diamond deposits is an iterative process of
analysis and interpretation, seeking to both identify and quantify
critical aspects of risk and uncertainty to acceptable levels as
defined in Yates and Hodson (2006). The major elements to
resource modelling at EKATI include diamond drilling for pipe
volume and internal geology which often occurs contemp-
oraneously with multiple phases of sampling for grade distribution
and diamond valuation. This enables both the deterministic and
geostatistical modelling of kimblerlite volume and domains, and
Advances in Orebody Modelling and Strategic Mine Planning I
 INTEGRATED STRATEGIC PLANNING AT EKATI DIAMOND MINE

Strategic
Pillars Strategic initiatives
imperative

Right 1. Relentless pursuit of Zero Harm to People

People,
Right Place 2. Employer of Choice

Maximize
Maximise 3. Relentless pursuit of Zero Harm to the environment
Licence to
operate
EKATI’s full 4. Operator of choice for communities and government

potential

5. Identify and optimize growth options

World Class
Asset
6. Creation of a cost conscious culture

7. Improve operating discipline

FIG 2 - EKATI strategy tree.

the simulation of diamond grade and calculation of resource
ranges at various mining scales. The classification of the resource
uses a combination of probabilistic and deterministic methods. It
is important to understand the internal geology of the kimberlite
pipe before undertaking any modelling exercise as internal phases
may act as either hard or soft boundaries for modelling. Trends
within and across domains may also play a key role due to both
the volcanic and sedimentary process involved in kimberlite
emplacement. Harrison et al, 2009 discuss the creation of
uncertainty based spatial models at EKATI as shown in Figure 3
using sequential Gaussian simulation (SGS) methods detailed in
Goovaerts (1997).
Each resource has a unique ore revenue (US$/tonne) which is
generated from the average recovered grade (carats per tonne)
multiplied by average diamond value (US$/carat). Average
diamond value (diamond reference value) is estimated for each
pipe or some cases multiple geological domains within a pipe
using exploration or production parcels ranging in size from
several hundred carats to tens of thousands of carats. The diamond
value estimate is relatively complicated compared to most
commodities. Diamonds occur in a skewed distribution of sizes
and qualities with a price variation up to three orders of magnitude
for a single diamond size. Diamonds within a kimberlite range in
quality from very low value boart (fibrous diamond) to very high
gem quality stones. More than 2200 categories comprise the
current EKATI diamond price book (PB). Average diamond value
is a function of diamond size distribution and diamond quality.
The highest value populations have both a coarse size distribution
and high proportion of high quality stones. The P50 diamond
reference value is a function of diamond size distribution, diamond
quality classification per size category (price point) and process
plant recovery. Uncertainty associated with diamond value
estimation is related directly to parcel size assuming a detailed,
well constrained value sort. This relationship is presented in
Figure 4, along with an indicative cost associated with bulk
sampling at EKATI. The uncertainty due to parcel size can be
estimated using Monte Carlo random testing of very large
production parcels. The ideal parcel size for commercial
kimberlite evaluation is approximately 5000 carats. However,
there are cost and time constraints in obtaining very large diamond
parcels at EKATI due to the difficult setting of the pipes (typically

FIG 3 - Sequential Gaussian Simulation (SGS) estimation of

3
carats/m , Jay Pipe, EKATI Diamond Mine (Harrison et al, 2009). FIG 4 - Parcel size, uncertainty and indicative sampling cost.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 145
B COUTTS
situated in lakes with less than ten metres of glacial overburden).
The evaluation is sequential due to multiple pipe development and
seasonal constraints (winter bulk sampling programs). At EKATI
those pipes in production have very large parcels resulting from
large scale production tests and the corresponding diamond value
estimates are very well constrained. Pipes still in the process of
investment evaluation studies have smaller, variably sized, parcels
with proportionate value uncertainty. The revenue impact of parcel
size risk can be positive or negative. The aim of increasing the size
of the diamond parcel through collection and processing of bulk
samples is to reduce the variability of the range of possible
diamond values. The valuation of the diamond parcels is
periodically re-fixed based on the prices achieved from the market
and this ‘price’ is called the price book. Prices are regularly
reviewed but for planning and market forecast purposes the Price
Book is fixed and the commodity price forecast (Diamond Price
Index) is applied to the fixed book price. EKATI’s product is not
sold continuously but in discrete events termed Diamond Sales.
Each time the price book is updated it is numbered according to
the diamond sales cycle to which it relates.
EKATI’s RDP is all inclusive, evaluating the operations full
risked mineral inventory which includes measured, indicated and
inferred resources as well as a full range of exploration, regional
and near-mine potential resources. The application of resource
range analysis generates an estimate of potential ranges within
the risked mineral inventory, and assigns probability weightings
to the ranges. At EKATI these ranges often incorporate potential
geologic phase based grade changes, and volumetrically minor
satellite kimberlites. The issue of technical certainty is managed
through planning cases built by the incremental addition of
resources based on the status of specific resource knowledge
and/or design maturity.
Engineering
Resource schedules are applied in the RDP in a range of
formats and obtained from a range of sources, including
underground production scheduled at a fixed rate in all cases as
provided by the underground operations or relevant engineering
studies, reserves on a bench by bench basis from designed pits,
scheduled tonnes on a bench by bench basis from Whittle shells
and Stockpiles. Schedules are updated as required to reflect
either current Whittle optimisations or the re-design of
operational mines.
Additional operating parameter assumptions that are generated
for the RDP include haul cycle times, haulage costs, incremental
bench costs, sinking rates, load factors and equivalent standard
truck factors.
Processing
Diamonds occur as discrete particles within a kimberlite pipe and
must be recovered as intact entities to be of commercial value.
Given the low concentration of diamond in kimberlite ore – one to
two parts per 20 000 000 on average worldwide, one to two parts
per 10 000 000 at EKATI – large volumes of material must be
processed to obtain economical quantities of product (ie
diamonds). Mineral processing at EKATI is based on the unique
physical properties of diamond that allow for standard reduction
and recovery by means of three main unit operations which are the
crushing and scrubbing of kimberlite ore followed by primary
concentration by heavy medium separation and then secondary
concentration by X-ray sorting. All of these operations are
performed on site. Final diamond cleaning and sorting is
completed at an off-site sorting and valuation facility. Screens of
specific size which are applied during the first two operations
control the recovery rate of diamonds from the material flow.
There are specific relationships between the screen size, plant
throughput rates and the value of the recovered product.
Reconciliation and validation of this recovery relationship is
146 Spectrum Series Volume 17
verified though auditing of plant reject material. Incorporation
of the current and correct plant operating strategy is a crucial input
to the RDP. Assessment of variations to this key parameter
is a critical element to ensure optimisation of the screen-recovery
-value relationship for a specific ore blend.
Financial
Prior to any work being carried out in the RDP it is important to
define the costs used. A cost model combining both short-term
and long-term budgets is used as the baseline for planned future
performance with evaluation and redistribution of costs undertaken
to ensure the correct costs are applied. Operating and capital cost
estimates applied in the RDP are sourced from either budget
estimates or relevant engineering studies. Unit costs are
determined for open pit and underground mining and the process
plant. Recognising the granularity of the analysis, a distinction is
made within unit costs between variable and fixed costs. Fixed
general and administrative costs comprise a large component of
the EKATI cost model, reflecting the expense of maintaining the
footprint required to operate in a remote and challenging
environment. All costs are in Canadian dollars on 100 per cent
terms and are exchanged to US dollars (US$) to be comparable to
revenue which is in US$. The active engagement of EKATI’s
management team is fundamental to achieving consensus in the
accuracy of costs generated and applied. The value impact of cost
variations is an important evaluation within the RDP. Range
analysis will be greatly facilitated by the development of an
integrated cost modelling system, discussed further in the
progression section.
Environmental
Reclamation and closure costs are sourced from a detailed closure
cost model. Working from the detailed cost estimating in the
model, pipe specific closure costs are allocated accordingly,
including those pipes where development has not yet commenced.
Where an underground operation is currently in production or
planned under an existing open pit, the open pit closure costs are
allocated to the underground operation. Pipe specific reclamation
costs are applied at the same time as pipe development costs. The
remainder (and far greater proportion) of the closure costs is
scheduled at the end of the mine life for each scenario. The timing
of the requirement to commit closure funds impacts the optimised
mining sequence determined and additional information in this
area will benefit future plans. Pipe specific ‘commencement of
production’ limitations are applied to reflect constraints associated
with project lead times resulting from environmental baseline
studies, permitting, regulatory approvals and engineering studies.
Marketing
Rough diamonds are not a commodity in the traditional sense
due to a lack of benchmark trading prices, a lack of homogeneity
in product characteristics/value, a lack of transparency in trading
due to proprietary producer-controlled marketing systems, and a
lack of transparency and homogeneity in value-add from rough
to polished to retail. Additionally, there is no established forward
market. The gem diamond market has historically been supply
constrained due to previous monopolistic controls, and demand
driven by emotional marketing. Since the commencement of
EKATI several key developments have resulted in a transition to
an open market characterised by a long-term scarcity of supply.
These developments include the emergence of multiple
competitive sources of supply, the abandonment of the market
custodian role by the incumbent and the liquidation of diamond
stockpiles. In addition, exploration has not succeeded in finding
substantial new resources while mature mines face declining
production. Demand is correlated with gross domestic product
(GDP). BHP Billiton’s diamond price index (DPI) is a
Advances in Orebody Modelling and Strategic Mine Planning I

long-range forecast in real terms of relative market value to a
fixed price book. It is applied to diamond valuations (US$/carat)
anchored to the same price book to forecast revenue in real US
dollars. The Diamond Price Protocol includes P99, P90, P50, P10
and P01 DPI forecasts that reflect the uncertainty range in
long-term price trends.
PROCESS
Tools
The EKATI Strategic planning process utilises a variety of
industry standard tools. These tools are summarised in this section,
with greater detail provided regarding the application of the
COMET optimisation software. Currently both short-term and
long-term scheduling is undertaken using a combination of
Microsoft Excel worksheets and Runge XPAC databases,
dependent upon the origin and application of the data. A project is
underway with the intent to transition all scheduling into XPAC.
Pit optimisation is performed using the Gemcom Whittle
optimisation tool. Pit shells are re-assessed as required dependent
upon changes in the key input values. Diamond values applied for
Whittle optimisation are those forecast to apply at the completion
of mining. Assessment of Whittle results for various economic
scenarios indicate that the P50 design captures almost all of both
P10 and P90 kimberlite tonnage. This apparent lack of sensitivity
to diamond value in pit optimisation reflects the discrete ore-waste
character of the kimberlite pipes. Given the relatively small size of
most of EKATI’s pipes, phase optimisation is only undertaken on
specific, larger resources. Pit design is completed utilising
Maptek’s Vulcan software suite.
Prior to 2003, the optimisation of EKATI’s value was
undertaken through Excel based Discounted Cash Flow models.
Strategic optimisation and options at EKATI are currently
evaluated through STOPS COMET software, both as a tool
within the RDP and for undertaking specific evaluations as
requested by management. The development and application of
an EKATI-specific COMET model enabled the objective
application and assessment of multiple policies, constraints and
options in a timely manner. COMET is an optimisation program
that is used to run alternative life-of-mine scenarios in a
relatively short period of time. It is a broad brush analysis tool
which is useful for making informed decisions on which options
warrant more detailed consideration. COMET is a software
package that iterates a mining sequence within a set of
predetermined constraints to give a resultant sequence which
generates the maximum Net Present Value (NPV). If a
Input Workbook
Data: Physical Resources
Phase File Generation
Workbook
Conversion
Data: Costs, Assumptions
Functions: Case Setup and
conversion of inputs into txt files
FIG 5 - Schematic of EKATI COMET model.
Advances in Orebody Modelling and Strategic Mine Planning I
INTEGRATED STRATEGIC PLANNING AT EKATI DIAMOND MINE
permissible option (eg an open pit) is not scheduled by COMET
this is an indication that this option is not value additive under
the applied conditions. For a fuller discussion on the algorithms
and assumptions applied within COMET the reader is referred to
Wooller (2007) and additional discussion on the integrated
applications of the software is provided by King (2007).
The COMET model currently applied at EKATI is based on
three Excel workbooks to allow the generation and control of both
input and output in a useable form. Figure 5 is a high level
overview of the COMET model. Specific elements within the
Inputs workbook include unique names for each resource and
design, ensuring no duplication. Sequencing and naming of
resources and designs is critical, as each worksheet contains
substantial resource and design specific data referenced
downstream in COMET. Data within each worksheet includes ore
tonnes and grade by phase, waste tonnes by rock type, dilution,
recovery and moisture factors, truck productivity and mining rates.
As the COMET optimisation process applies an identical
weighting and confidence to the evaluation of all individual
resources, resource uncertainty is managed through the
incremental assessment and evaluation of specific resource
packages of similar confidence. The Phase Generation File
workbook converts resource, physical and financial inputs into
format required by COMET. The scenario control worksheet
enables macro driven generation of required files. The workbook
populates phase (resources) and schedule (physicals and
financials) folders. Data within the workbook includes costs
(operating, capital and closure), processing scenarios and
economic assumptions. The COMET GUI contains specified
constraints, policies and schedules to enable the requested
optimisation to occur. The GUI contains the Optimisation Engine,
a macro for multiple scenarios and other control features. It
generates the physical and financial outputs for validation and
reporting. The resultant NPV is reported on a 100 per cent real
pretax $US in country basis. To enable further validation,
evaluation and reporting, the results of all COMET runs are
processed through a macro driven workbook within Pivot tables.
The nature of the optimisation process and applied constraints can
result in operational schedules which appear impractical as shown
in Figure 6. This is acceptable as the purpose of COMET is to
optimise the order of extraction. Further work on the maximised
NPV schedule during the next stage in the planning cycle
generates a deliverable mining schedule. Additional strategic and
operational applications of COMET at EKATI include both the
assessment of constraint specific ‘trigger values’ for options and
projects, and the evaluation of the impacts of dilution (in situ and
mining) and cost benefits of control programs.
Comet GUI Workbook
Data: Constraints, Policies, Schedules
Functions: Control Feature,
Optimization Engine and Output
Generated .txt Phase Files
Data: Physical Resources,
Costs, Assumptions
Spectrum Series Volume 17 147
B COUTTS
50
45
40
35
MILLIONS TONNES
30
25
20
15
10
0
20 11
2008
2009
2010
2012
Risks and opportunities
The core building block within the EKATI strategic planning
process is the present state, or the base case. Constraints applied to
the base case are the current and scheduled physical, operational,
economic and directional parameters. It also includes projects in
active execution. Considerable verification and focus is applied to
ensure the base case is fully and accurately represented and
modelled within the process tools. At EKATI the resultant NPV
from COMET evaluation of the base case is set at 100 per cent. The
results from all additional strategic assessment are always reported
as a percentage relative to the base case. The establishment of the
base case and the functionality of the COMET tool permit the
incremental assessment of development projects, tactical and
strategic options and alternate scenarios. Contingency and
sensitivity assessments facilitate the evaluation and understanding of
ranges (P99 through to P01) and certainties, upon both the base case
and further incremental scenarios. Variation of single constraints
within the model enables the identification and quantification of key
value drivers, opportunities for improvement and potential risks to
project value. The incremental value assessment of development
projects enables the relative prioritisation of required investment
studies. The quantification of the inherent certainty within each
development project provides an understanding of the associated
risk. As a progression of this, the variation of single constraints for
specific development projects enables the identification and
quantification of key project drivers such as operating costs, capital
costs or diamond value. The progressive building of project
groupings also permits an understanding of the impact to EKATI’s
value from changes to the project development schedules, again
facilitating the focus of internal energy upon the required projects.
Critical elements of option assessment include identifying the
key drivers for the preservation of options, and the trigger values
that will result in the activation of those options. At EKATI,
options considered and preserved include several large tonnage,
low margin resources which despite being un-economic under
current conditions are EKATI’s key drivers of scale. The trigger
values for these options are now the major goal of dedicated
Business Excellence and R&D programs. Plant expansions and the
resultant opportunity to reduce overall unit costs are also assessed.
148
TOTAL MATERIAL MOVEMENT
2 015
2 019
202 0
202 4
20 28
2013
2014
2016
2017
2018
2021
2022
2023
2025
2026
2027
2029
2030
FIG 6 - Representative COMET schedule.
That COMET routinely selects this option when it is made
available highlights the potential of a plant expansion to add value.
Greater certainty is required on costs and throughput rates for the
plant expansion option. A hypothetical exploration pipe has been
previously included to evaluate the scheduling and value impacts
of an exploration discovery. This emphasised the significant value
to EKATI of ongoing exploration aimed at such a discovery.
At EKATI both the base case and selected development options
are stress tested for their sensitivity to a range of potential
economic and operational scenarios. This includes an assessment
on any resultant changes to resource scheduling or developmental
priorities. These scenarios (either singularly or as combinations)
include high/low risk production applying a variance in process
plant throughput and reflecting poor winter road seasons or other
operational restrictions. Movements in DPI and value are assessed,
potentially caused either by new mines not coming on line or
substantially weakened global demand. Potential variance in the
scheduled production rates of high value ore sources, reflecting
either strong operational performance or operational restrictions, is
evaluated. Ranges in cost profiles, reflecting either success in cost
reduction programs or increase in direct or indirect costs (fuel, etc)
are included in the scenarios. The potential application of mining
methods alternative to conventional drill and blast for production
scale mining at specific kimberlite pipes is assessed. In addition,
the impacts of changes in the processing plant operational
parameters of throughput, screen size and value recovery,
including any required capital spend is evaluated.
DELIVERY
Effective communication and integration of value driven
strategic planning results within the BHP Billiton planning cycle
is critical to maximise the value of EKATI and develop
opportunities for a safe, long-term sustainable business.
Integration
Evaluation of EKATI’s base case, risks and opportunities
provides a series of quantified metrics that assist management
to determine the appropriate ‘go-forward’ strategic planning
Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I

options. The key metric used is the relative NPV. Other metrics
considered include the capital efficiency ratio (CER), the life of
the mining operation, total scheduled tonnes and the technical
uncertainty of development projects and options. A review of all
evaluations is held with the EKATI management team with the
purpose of verifying the results of the RDP process and
providing formal sign off on the go forward case(s) for the life of
asset (LOA) scheduling. Included in the recommendation for the
LOA schedules is which schedule will become the basis for the
5YP. The recommended ‘go-forward’ planning cases must
comply with specific JORC related and technical certainty
criteria. Two specific planning cases are recommended for
further assessment in the LOA. These are the LOA ‘Optimised
Without’ – the Base Case with projects and options in execution,
and the LOA ‘Optimised With’ – the above case with additional
value positive development projects and options. The go forward
cases from the RDP are approved by the CSG and therefore form
the basis of the LOA, 5YP and ultimately the 2 Year Budgets.
The go forward cases also lay the groundwork for the Portfolio
Valuation Review (PVR), which is the formal internal valuation
of the Asset. COMET provides an overall sequence through the
development options to maximise value. This sequence is used as
the basis for the PVR. These cases are scheduled as part of the
LOA process in XPAC for consistency of approach.
The planning cycle is an annual process, reflecting the reality of
operating and developing under constantly changing and evolving
physical, financial, regulatory and strategic parameters. At EKATI
major changes in both short and long-term planning has been
driven by variations in the certainty of resource elements such as
tonnage, grade and diamond valuation as a result of continued
sampling. Additional changes reflect the increasing understanding
of the key drivers of plant throughput and recovery. Ongoing
engineering studies result in changes to both the scale and
certainty of operating and capital cost estimates for development
projects. Additional planning cases are reviewed and approved by
management for assessment in the LOA, both ensuring the
preservation of options and to enable the most complete long
range view of the development potential of EKATI.
Communication
The integration of the RDP ensures formal communication of the
principal strategic planning case into EKATI’s business plan.
During all reviews significant use is made of graphical
presentation of data and results, including tornado, waterfall and
bubble charts. A review of all evaluations is held with both
the EKATI management team and the EKATI operational
leadership team to discuss the results of the RDP process.
Recommendations, with supporting justification, are provided
based on assessments undertaken during the process. These
recommendations include the confirmation of EKATI’s key
business drivers and the required trigger values to enable future
options. They address operational improvements, such as
changes to the processing configuration or confirmation of the
economic viability of alternate mining methods. Timelines for
permitting and/or environmental base line studies are
recommended to ensure deep preparation for supporting project
development schedules. The requirements for additional drilling
of development projects to further manage certainty of key
project specific variables such as grade, volume, hydrology or
structure is confirmed. Modifications are recommended to
EKATI’s mobile or fixed plant, such as future mobile fleet
requirements or changes in power generation to support required
infrastructure. Predictions of manning are generated based on the
future operational profile. The potential impacts for fine scale pit
design to increase reserves is highlighted and recommend future
specific design review work, including the assessment of split
shell designs in larger pits. The sensitivity on plant throughput
and pipe specific margins is provided to support the case for
development of both throughput data in resource models and
grade control programs. Specific elements of the EKATI
Advances in Orebody Modelling and Strategic Mine Planning I
INTEGRATED STRATEGIC PLANNING AT EKATI DIAMOND MINE
strategic plan are selectively shared with the wider stakeholder
community, to both emphasis development opportunities for
EKATI as a safe, long-term sustainable business and to seek to
proactively engage community expectations.
PROGRESSION
Over the preceding 13 years strategic planning at EKATI has
constantly evolved. The current process delivers value focused
integration throughout the planning cycle, from directional to
delivery. Future improvements to the process, some currently in
execution, are intended to proactively progress this ongoing
evolution.
Inputs
A potential development is the imposition of a carbon tax in
Canada. Given EKATI’s current full dependence upon diesel for
supply, heat, fleet and power, this tax would have a significant
impact upon both the base case and development projects. The
sensitivity assessment of this on selected planning cases would
quantify both the costs and benefits of potential mitigation
activities such as the development of alternate energy sources. A
significant drawback in the current process methodology is the
de-facto imposition of annual production targets as constraints,
through the ‘fixing’ of resource development to honour
scheduled tonnes. This removes the ability of the optimisation
software to generate a truly NPV maximised result, and
unconstrained analysis is critical to evaluate and understand key
NPV drivers. Given that EKATI is a mill constrained operation,
the assessment of potential stockpiling strategies is warranted.
This has not been fully evaluated to date, in part due to the
historical acceptance of production targets as constraints.
EKATI’s unique geographical and physical parameters have
required innovative solutions. Current R&D programs into
alternate methods of mining, ore haulage, transportation and
energy continue this tradition. The integration of R&D into
strategic planning enables both evaluation and prioritisation of
opportunities. Future large scale development projects at EKATI
are low margin. One of the solutions to this challenge is the
selective sequencing and mining of higher grade zones within
these resources. Evaluation of the resultant impacts may provide
justification for the development of required supporting
programs such as grade control.
Process
An integrated planning system currently being developed is
based on an XPAC-XERAS scheduling and budgeting core with
programmed links to COMET and Whittle optimisation. It will
facilitate both the consistent secure provision of input values and
the iterative assessment of development options. The system will
enable alignment between the many suppliers and customers of
the contained data. The development of pipe-specific throughput
factors will enable the application of time factored cash flow
grades ($/hour) at EKATI, further enhancing both directional and
deliverable planning and scheduling. Given that specific
elements within EKATI’s planning process contain inherent
highly variable uncertainties, the application of real options
Analysis (ROA) to both increase the explicit understanding of
underlying assumptions, and to better predict and understand the
value of project flexibility must be assessed. A real option is the
right to undertake a business decision and ROA applies put
option and call option valuation techniques to these decisions.
Delivery
It is an ongoing obligation of strategic planning at EKATI (and
elsewhere) to challenge accepted paradigms. Engagement in such
discussion is most effectively achieved by ‘Size of the Prize’
Spectrum Series Volume 17 149
B COUTTS
assessment of the potential options. For example, what are the
resultant increases in both project value and life from achieving a
specific and sustainable unit operating cost? A key element in
EKATI’s long-term sustainability is proactive engagement of
regulators and the wider community in the development of fully
detailed and cost inclusive Closure Plans. The integration and
evaluation of closure within strategic planning is critical.
CONCLUSIONS
Effective strategic planning at EKATI requires both the
engagement of the business into the process, and the integration of
the results into deliverable schedules and developments. This is
achieved through an inclusive review of strategic direction and an
annual, comprehensive, inter-linked planning system to both
optimise and challenge the present state. Integrated strategic
planning will be a key element in successfully realising EKATI’s
potential as the lowest cost premier producer of diamonds in the
developed world.
ACKNOWLEDGEMENT
BHP Billiton Diamonds Inc management are thanked for the
permission to publish this paper. This paper presents a selective
snapshot of the continuous process of strategic planning at
EKATI and the contribution of all BHP Billiton staff and
consultants is most warmly recognised.
REFERENCES
BHP Billiton, 2008. Practioner’s guide to planning.
BHP Diamonds Inc, 1996. NWT diamonds project feasibility report.
150 Spectrum Series Volume 17
Dyck, D R, Oshust, P O, Carlson, J A, Nowicki, T E and Mullins, M P,
2004. Effective resource estimates for primary diamond deposits
from the EKATI Diamond Mine, Canada, in Proceedings Eighth
International Kimberlite Conference Selected Papers (eds: R H
Mitchell, H S Gruetter, L M Heaman, B H Scott Smith and
T Stachel), vol 1, pp 317-335.
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
483 p (Oxford University Press: New York).
Harrison, S, Leuangthong, O, Crawford, B and Oshust, P, 2009.
Uncertainty based grade modelling of kimberlite: A case study of the
Jay kimberlite pipe, EKATI Diamond Mine, Canada, in Proceedings
Ninth International Kimberlite Conference.
King, B, 2007. Integrated strategy optimisation for complex operations,
in Orebody Modelling and Strategic Mine Planning, second edition
(ed: R Dimitrakopoulos), pp 31-35 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Kleingeld, W J and Nicholas, G D, 2007. Diamond resources and
reserves – Technical uncertainties affecting their estimation,
classification and valuation, in Orebody Modelling and Strategic
Mine Planning, second edition (ed: R Dimitrakopoulos), pp 207-213
(The Australasian Institute of Mining and Metallurgy: Melbourne).
Wooller, R, 2007. Optimising multiple operating policies for exploiting
complex resources – An overview of the COMET scheduler, in
Orebody Modelling and Strategic Mine Planning, second edition (ed:
R Dimitrakopoulos), pp 289-296 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Yeates, G and Hodson, D, 2006. Resource classification – Keeping the
end in sight, in Proceedings Sixth International Mining Geology
Conference, pp 97-104 (The Australasian Institute of Mining and
Metallurgy: Melbourne).
Advances in Orebody Modelling and Strategic Mine Planning I

Stochastic Long-Term Production Scheduling of Iron Ore
Deposits – Integrating Joint Multi-Element Geological
Uncertainty
J Benndorf1 and R Dimitrakopoulos2
ABSTRACT
Meeting production targets in terms of ore quantity and quality is critical
for a successful mining operation. In situ grade variability and uncertainty
about the spatial distribution of ore and quality parameter cause both
deviations from production targets and general financial deficits. A
stochastic integer programming formulation (SIP) is developed herein to
integrate geological uncertainty described by sets of equally possible
scenarios of the unknown orebody. The SIP formulation accounts not only
for discounted cash flows and deviations from production targets, but also
discounts geological risk, while accounting for practical mining.
Application at an iron ore deposit in Western Australia shows the ability of
the approach to control risk of deviating from production targets over time.
Comparison shows that the stochastically generated mine plan exhibits less
risk in deviating from quality targets than the traditional mine planning
approach based on a single interpolated orebody model.
INTRODUCTION
Long-term mine planning and production scheduling aim to define
the ‘best’ mine plan subject to the constraints imposed by physical
and geological conditions, policies and the operational mining
approach. The term ‘best’ is defined by management objectives.
These typically include maximising the monetary value of the
mining project as well as meeting customer expectations and
guaranteeing a safe operation. The expectations of customers are
defined largely in terms of ore tonnage and ore quality
characteristics to be delivered. In the case of multi-element
deposits, ore quality characteristics are defined by multiple inter-
correlated elements. For example, in iron ore deposits, the
elements iron (Fe), phosphorus (P), silica (SiO2), alumina (Al2O3)
and loss of ignition (LOI) are critical for ore quality. Additionally,
in many cases ore is produced out of multiple pits with different
ore characteristics. The goal of any global, long-term mine
planning approach is to send the most homogeneous ore blend out
of multiple pits, meeting customer specifications, while
guaranteeing optimal pit development and maximising the utilis-
ation of available mineral resources. In practice, however, when
implementing a mine plan, differences frequently occur between
the produced ore quantity and quality characteristics. It is well
recognised that uncertainty in the description of the spatial
distribution of grades of various pertinent elements in the orebody
as well as their in situ variability are major contributors to these
differences.
Traditional approaches to mine planning optimisation are based
on a single estimated model of the orebody that is unable to
account for in situ variability and uncertainty associated with the
description of the orebody (David, 1977; David, 1988). Contrary
to estimation techniques, a different set of techniques provide a
tool to address shortcomings of estimation methods, termed
conditional simulation (Goovaerts, 1997; Chiles and Delfiner,
1. GAusIMM, MIBRAG MBH, Germany.
Email: JoergBenndorf@gmx.de
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montreal QC H3A 2A7, Canada.
Email: roussos.dimitrakopoulos@mcgill.ca
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
1999; Dimitrakopoulos, 2007). Based on drill hole data and their
statistical properties, conditional simulations generate several
equally probable models (or scenarios) of a deposit, each
reproducing available data and information, statistics and spatial
continuity, that is, the in situ variability of the data. The difference
between the equally probably scenarios are a quantitative
measure/description of uncertainty. The subsequent integration of
this grade uncertainty and local variability into mine planning
optimisation allows for the understanding and control of
geological risk. This in turn aims to decrease project risk and
increase profitability.
The detrimental effects to mine planning optimisation from
ignoring in situ grade variability and uncertainty in the
description of orebodies are well documented (Ravenscroft,
1992; Dowd 1997; Dimitrakopoulos, Farrelly and Godoy, 2002;
and others). For example, Dimitrakopoulos, Farrelly and Godoy
(2002) show the danger of relying on estimated (average type)
orebody models when optimising. In their example, net present
value (NPV) assessment of the conventionally generated
life-of-mine schedule using simulated scenarios of the orebody
shows the most likely NPV to be materialised standing at 25 per
cent lower than forecasted. The substantially positive
contribution of accounting for grade uncertainty through multiple
simulated scenarios and new stochastic optimisation approaches
is also well documented. Godoy and Dimitrakopoulos (2004)
show a long-term production scheduling approach based on
simulated annealing applied to a gold mine to result in a 28 per
cent increase of project value compared to the conventional
approach. Leite and Dimitrakopoulos (2007) show the same
order of improvement using this approach at a copper deposit. A
more general and flexible long-term production scheduling
method that allows the control of geological risk between
production periods in terms of magnitude and variability is based
on stochastic integer programming or SIP (Birge and Louveaux,
1997) and it is documented in Ramazan and Dimitrakopoulos
(2007b) and Leite and Dimitrakopoulos (2010, in this volume).
An application of the SIP formulation to the long-term
production scheduling of a single-element deposit demonstrates
its effectiveness and advantages in terms of additional project
value and associated risk management even for a relatively short
life-of-mine.
This paper contributes a mine planning optimisation approach
that addresses joint multi-element grade uncertainty, as common
in many mineral deposits, such as iron ore. More specifically, the
stochastic integer programming approach of Ramazan and
Dimitrakopoulos (2007b) is expanded to:
• multi-element deposits, and
• includes new mineability constraints to facilitate accessibility
and equipment size constraints.
In addition, the formulation developed herein is exhaustively
tested in an application at an open pit iron ore mine in Western
Australia and within the context of multi-pit production
planning. Testing includes the ability of the SIP to control the
risk of deviating from production targets in terms of ore quality
characteristics. In the next sections, the stochastic mathematical
programming formulation is first presented. The application and
testing of the formulation are presented.
Spectrum Series Volume 17 151
J BENNDORF and R DIMITRAKOPOULOS

STOCHASTIC PRODUCTION SCHEDULING
Global optimisation of long-term production scheduling
addresses issues of optimal sequencing considering multiple pits,
multiple elements, blending issues, stockpiling options and
alternative processing or product options (Whittle, 2007). The
task of long-term production scheduling in a multi-pit operation
can be divided into two stages. The first stage is a multi-pit
scheduling approach, which defines ultimate pit outlines as well
as proportions and element qualities, where each pit and period
contribute to the global target in order to optimise the global
asset. In the second stage the physical extraction sequence of
blocks in each single pit is defined as constraints to production
rates and targeted element grades implied by the multi-pit
scheduling approach. This contribution concentrates on the
long-term scheduling of a single pit – multi-pit scheduling
approaches have already been successfully implemented, eg
BLASOR, developed in BHP Billiton’s Technology group (Stone
et al, 2007).
The goal of long-term production scheduling under grade
uncertainty of single pits is to define a physical extraction
sequence of blocks over periods so as to meet multiple goals.
These goals include:
• best mine development and best use of available mineral
resources for a maximisation of the monetary value of the
asset,
• control of risk of deviating from production targets, and
• guarantees of a safe operation.
In this context, controlling the risk of deviating from production
targets is a major contribution and involves controlling
probabilities and magnitudes of deviations from production
targets, as well as fluctuation of produced grades over periods. The
underlying geological uncertainty is captured by a set of
conditionally simulated orebody models. Generally, production
targets may be in terms of produced ore and waste tonnes and
grades of different elements. Constraints are in terms of
practicality of the schedule guaranteeing equipment accessibility,
mining capacity, processing capacity, geotechnical aspects as well
as blending requirements.
P K
− ∑ ∑ (c SM .Y1tj )
t=1 j=1
where:
P is the number of periods
N denotes the total number of blocks to schedule
S represents the number of simulated orebody models used to
capture geological uncertainty
R is the number of targets including grade targets for different
elements and ore tonnage targets
c ti represents the economic contribution of block number i
when mined in period t and is a representation of the
expected economic value over all values of block i at time t
derived from each realisation s E{(NPV)ti }
x ti is a variable representing the percentage of block i mined in
period t – if an x ti variable is defined as binary (0 or 1), it is
assigned 1 if block i is mined in period t and assigned 0 if
not
s
qu tr is the upper deviation from production target r at time t
considering orebody model s
yu tr is the unit cost of squrt to penalise excess production
s
ql tr is the lower deviation from production target r at time t
considering orebody model s
yl tr is the unit cost of sql tr to penalise a deficit in production
Y1tj is the number of surrounding blocks, which are not mined
in the period t or earlier when mining block j
Surrounding blocks are those, which are no more than three
blocks apart in each direction (Figure 1). The costs cSM are
penalties associated with Y1tj . Note that this penalty only applies
to a subset K of all blocks N. To avoid overlapping, only every
third block in each direction is considered to be the central block j.
The first part of the objective function is used for maximising
the discounted economic value in the context of the global
optimisation. Note the global multi-pit approach accounts for
interactions between different pits and aims to maximise usage of

Stochastic formulation for long-term production

scheduling

A general formulation for long-term production scheduling

under geological uncertainty for multi-element deposits based
on SIP is presented next. It is based on the single element
formulation in Ramazan and Dimitrakopoulos (2007b). The
objective function and relevant constraints are explained in
detail.
j
Objective function

The SIP objective function, presented here for scheduling

multi-element single deposits, combines several goals. It aims to
generate a production schedule that optimises the economic pit
development considering constraints imposed by the global
multi-pit approach, while minimising deviations from production
targets in terms of tonnages and ore-quality as well as minimising
costs of non-smooth mining. Equation 1 presents the three parts of
the objective function:
P N S P R
Maximise ∑ ∑ c ti .x ti -∑ ∑ ∑ ( s qu tr .yu tr + ql tr .yl tr ) (1) FIG 1 - Inner and outer window around block j in smooth mining
t = 1 i= 1 s = 1 t = 1 r= 1 constraint (after Dimitrakopoulos and Ramazan, 2004).

152 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 STOCHASTIC LONG-TERM PRODUCTION SCHEDULING OF IRON ORE DEPOSITS
resources and global value. The first part in Equation 1 maximises
the local NPV of the single pit under consideration aiming to
define an optimal mine development constricted by the global
plan. It accounts for profit-defining aspects, such as stripping ratio.
The discounted economic block value is calculated as expected
value from each realisation. The second part of the objective
function handles the deviations from production targets imposed
by the multi-pit scheduling approach for each simulated orebody
model s including grades of all elements and ore tonnage. By
optimising over S possible scenarios, captured through multiple
equally probable orebody models, this part of the objective
function aims to control uncertainty and variability of the
produced grades and ore tonnage. The magnitude of grade
variability in the generated schedule is controlled for each element
e considered and time period t by penalties associated with
deviations s ql tr and s qu tr . Note that deviations for each target and
period yu tr and yl tr are calculated by the corresponding constraints,
which are the grade constraint and the ore tonnage constraint. Part
three of the objective function controls smooth mining by
penalising not mining adjacent blocks in same period, the central
block j is scheduled, or earlier (Figure 1). Yl tj represents hereby
the percentage of the eight directly adjacent blocks and the 25
blocks that are two block-widths distant, which have not been
mined in the same period as block j. Deviations of smooth mining
for each considered block j and period t Yl tj are calculated in the
smooth mining constraint. The priorities of the three competing
parts in the objective function are controlled by the magnitude of
corresponding cost parameters for each part relative to each other.
The mine planner has to adjust these parameters so to define the
best schedule that compromises his objectives, for example the
level of risk the planner is willing to accept. The related
constraints are detailed by Benndorf (2005).
Controlling risk over time for different objectives
As presented in the previous section, penalties associated with
deviating from production targets introduced in the objective
function aim to control risk of deviation for each element. These
penalties can be defined in different magnitudes for each element
and period. This enables the mine planner to control the risk for
each element over time. The ability to control the risk over time is
a concept introduced by Dimitrakopoulos and Ramazan (2004)
using a geological risk discount rate. This discount rate is directly
applied to penalties and thus controls the risk distribution between
periods. A high geological discount rate indicates that the SIP
formulation herein is emphasised to generate a schedule that is
less risky in early periods than in later periods. This may be useful
when the operation aims to mine less risky parts of the deposits in
early periods and more uncertain parts in later periods. As mining
progresses, more information about those uncertain parts will
become available in form of operational exploration. A geological
discount rate of zero per cent generates schedules that are
expected to exhibit a similar level of risk in all periods. The
difference between penalties applied to upper deviations and lower
deviations defines the priority of upper and lower deviations from
targets. For example, it may be more important in an operation to
keep the deficit in production as low as possible while excess
production may not be of importance.
PRODUCTION SCHEDULING UNDER
UNCERTAINTY – AN APPLICATION AT
YANDI CENTRAL 1 IRON ORE DEPOSIT,
WESTERN AUSTRALIA
Next, mine production scheduling under multi-element grade
uncertainty is applied to the Yandi Central 1 iron ore deposit in
Western Australia. The first part describes the Yandi Central 1
deposit focusing on geology, mining operation and current
production scheduling practice. The problem specification and
Advances in Orebody Modelling and Strategic Mine Planning I
description of input data are discussed subsequently, in particular
the process of incorporating the stochastic production scheduling
approach of a single deposit into the global multi-pit scheduling
problem. The input in terms of simulated orebody models is
presented as well as the operational, economical and risk
controlling parameters. Following, the practical approach of
scheduling Yandi Central 1 is detailed, including the practical
implementation of the scheduling formulation and the manual
mine design to convert results to a practical schedule. A
comparison between schedules generated using a stochastic
formulation to those using a deterministic formulation
considering one estimated orebody model is found at the end of
this section and demonstrates the benefit of the stochastic
approach.
Yandi operation and current production
scheduling practice
The Yandi Central 1 deposit is part of the larger Yandi channel
iron deposits (CID), which occurs alongside the Marillana-
Yandicoognica Creek system about 120 km northwest of
Newman, Western Australia. This deposit is part of the Yandi
joint venture operation, which includes multiple pits. The
fundamental objective of this complex operation is the achieve-
ment of customer defined on-grade shipments at lowest costs by
optimally blending from different pits with a diverse range of
resource grades. Critical geochemical parameters when evaluating
the deposit are iron content (Fe), silica content (SiO2), alumina
content (Al2O3), phosphorus content (P) and the water and organic
content measured as loss on ignition (LOI), as they influence the
physical and chemical properties of the product and the
performance of the beneficiation process.
For the global multi-pit optimisation of the Yandi joint venture
operation, BHP Billiton’s Technology group developed a
scheduling-algorithm, termed BLASOR (Stone et al, 2007).
Among other details BLASOR assigns targets in terms of
produced ore tonnes and grades for each period to each pit as
contributing to the global target. Although BLASOR, as used
here, accounts for multiple elements, the approach is based on a
single estimated orebody model and does not incorporate local
uncertainty and in situ variability.
Problem specifications and input for scheduling
The in situ variability and the incomplete knowledge of the
spatial distribution of the elements in the orebody are most
critical for meeting customer specifications. In order to
incorporate in situ variability and uncertainty of geochemical
parameters in mine production scheduling, techniques for
optimisation under uncertainty can be employed. The application
of stochastic mine production scheduling to Yandi Central 1 is
based on stochastically simulated orebody models generated
using the computationally joint direct block simulation approach
(Boucher and Dimitrakopoulos, 2009). Operational, economic
and risk defining parameters are explained in subsequent sections
in more detail.
Stochastic orebody models at Yandi Central 1
The basis for mine production scheduling under geological
uncertainty is a series of simulated orebody models of the
deposit. For this case study, 20 simulated orebody models of the
main ore zone (MOZ) are used, generated by Boucher (2003)
using the joint block simulation technique detailed in Boucher
and Dimitrakopoulos (2009). This joint-simulation of the five
considered elements Fe, P, SiO2, Al2O3 and LOI guarantees the
local reproduction of cross-correlation between the elements.
Note that Fe is strongly correlated with the elements SiO2 and
Al2O3. Each of the resulting orebody models contains 3049
Spectrum Series Volume 17 153
J BENNDORF and R DIMITRAKOPOULOS

-86500 N Lower bench

-86000 N

-85500 N

-13000 E

-15000 E
-11000 E

-86500 N Middle bench

-86000 N

Fe-content in %
-85500 N
-13000 E

-15000 E
-11000 E

0.0-57.4
57.4-58.2
58.2-58.7
58.7-60.6
Upper bench
-86500 N

-86000 N

-85500 N
-13000 E

-15000 E
-11000 E

FIG 2 - Spatial distribution of Fe-grades in realisation number five for the lower, middle and upper bench.

blocks in total. Block dimensions are 25 m × 25 m × 12 m,
representing typical mining units. Each block contains the
attributes total tonnage, ore tonnage as well as total content of
each element Fe, P, SiO2, Al2O3 and LOI. As an example, a map
of the spatial distribution of Fe grades in the orebody model is
presented in Figure 2 for the case of simulated realisation
number five.
Operational parameters
Operational parameters, including ore production and required
qualities are defined by the global multi-pit scheduling approach
undertaken by BLASOR. BHP Billiton Iron Ore provided
scheduling results defining the contribution of Yandi Central 1 to
the global target for the following five years referred to as periods.
For confidentiality reasons, BLASOR results are scaled (Table 1).
Ideally, shipping grades are to be delivered with nearly zero
variability. Since this is unlikely, the industry sets target bands
limited by an upper and lower bound. Grades should not fall
outside this band. Table 1 summarises initial ore tonnage and
grade limits. The differentiation between ore and waste prior to
the optimisation is realised through an Fe grade cut-off of
0.56 per cent. Further, it is assumed that the operation is flexible
enough to account for different ore and waste production rates
between periods. For this reason, the maximum mining capacity,
including ore and waste production, was set to 20 000 000 t,
which is about 5 000 000 t more than the maximum rate. Due to
the flat geometry of the deposit, one slope region is sufficient to
characterise the geotechnical constraints. The general slope angle
is set at 45°.
Economic and risk-controlling parameters
Table 2 presents the economic parameters, including price,
mining and processing costs and discount rates. Mining costs
include blasting, extraction and transportation costs – processing
costs account for crushing, conveying and stockpiling. Two
discount rates are identified, the economical discount rate and
the geological discount rate. The economical discount rate
discounts cash flows over periods, while the geological discount
rate controls the risk of producing grades that fall outside the
limits over the periods. Recovery is 100 per cent. The stochastic
scheduling approach applied in this case study is concerned with

TABLE 1
Ore tonnage and grade constraints for scheduling Yandi Central 1.

BLASOR scheduling results of Yansi Central 1 for first periods

Page number Ore tonnage (wt) Fe (%) P (%) SiO2 (%) Al2O3 (%) LOI (%)
1 14 000 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0
2 10 000 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0
3 10 000 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0
4 9 000 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0
5 7 200 000 57.1 - 59.4 0.032 - 0.038 4.6 - 5.2 0.90 - 1.05 9.5 - 11.0

Note: Ore/waste cut-off grade is Fe> = 56 per cent.

154 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 STOCHASTIC LONG-TERM PRODUCTION SCHEDULING OF IRON ORE DEPOSITS

define the pit design for each period and provide a mineable
TABLE 2 production schedule. Parameters used in this designing process
Economical parameters for long-term production scheduling of the are a 12 m bench height, 45° slope angle and a 5 m berm
Yandi Central 1 iron ore operation.
between two toe and crest string, a road width of 25 m and a
Parameter Cost/price eight per cent ramp incline. The lower part of Figure 3 shows a
south-east isometric view of the resulting smooth schedule.
Price per ton recovered metal $30
Benndorf (2005) demonstrated that this type of smoothing has no
Mining costs per ton $5 significant impact on the results, which means that the smoothed
Processing costs per ton $5 schedule is still near to optimal.
Economical discount rate 10%
Geological discount rate 10% Evaluating results

In addition to produced ore and waste tonnage, results are

the risk of not meeting production targets of produced
element-grades. Penalties for deviating from production targets
are set initially to $1/unit of deviation.
THE PRACTICAL SCHEDULING APPROACH
Initial run and practical mine design
The upper part of Figure 3 shows results of an initial run using
above specified parameters. The extraction sequence appears
smooth and feasible, however there are few blocks scheduled
surrounded by blocks scheduled in different periods. To generate
a practical mining schedule that guarantees minimum mining
width and equipment accessibility, results of the stochastic
formulation are refined using manual mine design and haul road
construction. These standard tools are available in commonly
used mine scheduling software packages. In this study open pit
design from Earthworks Datamine is used (Data- mine manual,
2002). The schedule generated by the formulation can be used as
a guideline to construct polygons for each period and bench.
These polygons, in combination with haul roads and ramps,
Period 1
Period 2
0m Period 3
50 Period 4
Period 5
Period 1
Period 2
0m Period 3
50 Period 4
Period 5
FIG 3 - Stochastic schedule in ultimate pit – before (upper part)
and after (lower part) smoothing using manual design.
evaluated in terms of risk profiles of produced grades per period,
in particular for Fe, SiO2, Al2O3, P and LOI (Figure 4). For each
period the grades are shown considering each simulated orebody
realisation, which represent possible scenarios based on
information available. The spread of the different realisations
provide an indication about uncertainty in produced grades per
period when extracting the deposit according to the generated
schedule. Analysing the risk profiles of Fe, P and LOI results
concludes that there is no risk of deviating from production
targets. SiO2 and Al2O3 appear to be more critical in meeting
production targets. For example, four out of twenty simulated
orebody models for SiO2 indicate a deviation from the lower
target in period one. Thus, there exists a 20 per cent chance of
not meeting production targets for SiO2 in period one.
The ability to control risk
A major contribution of the presented scheduling formulation is
the ability to control risk of deviating from production targets
considering different quality parameters. As experienced in the
initial run, SiO2 and Al2O3 appear most critical in meeting
targets. To investigate the ability to decrease risk, three different
schedules were generated applying different penalties to both
critical elements. The three schedules were generated using low
(1$ per unit deviation per tonne), medium (10$ per unit deviation
per ton) and high penalties (100$ per unit deviation).
Figure 5 shows the extraction sequence of the lower bench for
each schedule. In the case of each schedule, the deposit would be
extracted in a different sequence. The dispersion of the schedules
increases with the magnitude of the penalties. In the case of low
penalties, the extraction sequence is smooth. Although medium
penalties generate a more dispersed schedule, it is still smooth
enough to be converted to a feasible schedule using manual mine
design. High penalties generate a very dispersed schedule, which
could hardly be efficiently realised. The dispersion is an
expression of a higher selectivity, necessary in order to produce a
homogeneous product in a tight quality band. Figure 6 shows the
risk profiles for SiO2 and Al2O3 for the three generated schedules.
In case of SiO2, the effect of increasing penalties already becomes
obvious in the case of medium value penalties. Compared to the
low penalty case, the fluctuation of grades between periods
decreases significantly and there exists only a slight probability of
deviating from targets in period 2, 3 and 4. Higher penalties
improve the result only marginally. In the case of Al2O3, a
decrease in probability of deviating from targets is recognisable
with higher penalties, however, there still exists a certain amount
of risk. This is an expression of a high in situ variability and
uncertainty of the element, which cannot be avoided by blending
in the pit. A solution here, to decrease the risk, could be to blend
the ore with ore from different mines, where Al2O3 is less variable
and uncertain.
Generally, this evaluation of the scheduling formulation
demonstrates that less risk of deviation comes with a cost of

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 155
J BENNDORF and R DIMITRAKOPOULOS

Production Yandi Central 1 Fe Grades

Limits: 57.1 % - to 59.4 %
16 59.50%
Millions

14
59.00%
12

10 58.50%

Grade in %
Ore
Tons

8 Waste
Total 58.00%
6

4
57.50%
2

0 57.00%
1 2 3 4 5 0 1 2 3 4 5 6

Period Period

Silica Grades Alumina Grades

Limits: 4.6 % to 5.2 % Limits: 0.90 % to 1.05 %
5.3% 1.05%

5.1% 1.00%
Grade in %

Grade in %
4.9% 0.95%

4.7% 0.90%

4.5% 0.85%
0 1 2 3 4 5 6 0 1 2 3 4 5 6
Period Period

LOI P-Content
Limits: 10.0 % to 11.0 % Limits: 0.032 % - to 0.038 %
10.70% 0.038%

0.037%
10.60%
Grade in %

Grade in %

0.036%
10.50%
0.035%

10.40% 0.034%

0.033%
10.30%
0 1 2 3 4 5 6 0.032%
Period 0 1 2 3 4 5 6
Period

FIG 4 - Results of stochastic scheduling in terms of ore and waste tonnages and risk profiles for Fe, SiO2, Al2O3, P and LOI.

Lower bench
Low penalties ( 1 per unit deviation)

Medium penalties (10 per unit deviation)

High penalties (100 per unit deviation)

Period 1
Period 2
Period 3
Period 4
Period 5

FIG 5 - Different extraction schedules depending on the magnitude of penalties for the lower bench.

156 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 STOCHASTIC LONG-TERM PRODUCTION SCHEDULING OF IRON ORE DEPOSITS

Alumina Grades
Limits: 0.90 % to 0.95 %
1.05%

Low penalties ( 1 per unit deviation)

1.00%
Grade (in %)

0.95%

Alumina Grades
Limits: 0.90 % to 0.95 %
0.90%
1.05%

0.85%
1.00%
0 1 2 3 4 5 6

Grade (in %)
Period
0.95%

Medium penalties (10 per unit deviation) 0.90%

0.85%
Alumina Grades
0 1 2 3 4 5 6
Limits: 0.90 % to 0.95 %
Period
1.05%

1.00%
Grade (in %)

0.95%

0.90%
High penalties (100 per unit deviation)

0.85% Limits used in SIP formulation

0 1 2 3 4 5 6
Period

Silica Grades
Limits: 4.7 % to 5.1 %
5.3%

Low penalties (1 per unit deviation)

5.1%
Grade (in %)

4.9%

Silica Grades
4.7% Limits: 4.7 % to 5.1 %
5.3%

4.5%
0 1 2 3 4 5 6 5.1%
Period
Grade (in %)

4.9%

Medium penalties (10 per unit deviation) 4.7%

4.5%

Silica Grades 0 1 2 3 4 5 6

Limits: 4.7 % to 5.1 % Period

5.3%

5.1%
Grade (in %)

4.9%

High penalties (100 per unit deviation)

4.7%

4.5% Limits used in SIP formulation

0 1 2 3 4 5 6
Period

FIG 6 - Risk profiles for produced grades (alumina and silica) depending on penalties.

higher selectivity, which is caused by the two competing
objectives in the objective function: minimise risk of deviating
from production targets, and generate a smooth schedule.
CONCLUSIONS
A new stochastic integer programming based mine production
scheduling approach, which considers jointly multi-element
geological uncertainty, is presented and successfully applied to
production scheduling at the Yandi Central 1 deposit, WA. It is
demonstrated that the SIP formulation presented, can be
implemented as part of a multi-pit scheduling approach. In this
application, results from BLASOR, a multi-pit scheduling
optimisation approach, are used to define the contribution of the
Yandi Central 1 deposit, Western Australia, to the global target per
period in terms of desired grades of elements and ore tonnages.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 157
J BENNDORF and R DIMITRAKOPOULOS
Results demonstrate the ability of the stochastic approach to
control risk of deviating from production targets for critical
quality defining elements. A comparison between the stochastic-
ally generated production schedule and a schedule generated
using one estimated orebody model illustrated the benefit,
stochastic models can generate. The stochastic schedule shows a
higher probability in meeting production targets, which decreases
overall project risk and can increase project value.
ACKNOWLEDGEMENTS
Funding was partially provided by NSERC Discovery Grant
239019-06 and BHP Billiton, Rio Tinto, AngloGold Ashanti and
Xstrata.
REFERENCES
Benndorf, J, 2005. Efficient sequential simulation methods with
implications on long-term production scheduling, MPhil thesis
(unpublished), The University of Queensland, Brisbane, 225 p.
Birge, J R and Louveaux, F, 1997. Introduction to Stochastic Pramming,
421 p (Springer-Verlag: New York).
Boucher, A, 2003. Conditional joint simulation of random fields on block
support, MPhil thesis (unpublished), University of Queensland,
Brisbane, 168 p.
Boucher, A and Dimitrakopoulos, R, 2009. Block-support simulation
of multiple correlated variables, Mathematical Geosciences,
41(2)142-158.
Chiles, J P and Delfiner, P, 1999. Geostatistics, Modelling Spatial
Uncertainty, 695 p (John Wiley and Sons: New York).
Datamine manual, 2002. Datamine-guide graphical mining software
reference manual.
David, M, 1977. Geostatistical Ore Reserve Estimation, 364 p (Elsevier:
Amsterdam).
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
Reserve Estimation, 216 p (Elsevier: Amsterdam).
Dimitrakopoulos, R, 2007. Risk analysis in ore reserves and mine
planning: Conditional simulation concepts and applications for the
mining industry, The Australasian Institute of Mining and Metallurgy
– McGill 2007 Professional Development Seminar Series, 385 p.
Dimitrakopoulos, R, Farrelly, C and Godoy, M, 2002. Moving forward
from traditional optimisation: Grade uncertainty and risk effects in
open pit design, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 111:A82-A87.
Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty based production
scheduling in open pit mining, SME Transactions, 316:106-112.
158 Spectrum Series Volume 17
Dowd, P A, 1997. Risk in minerals projects: Analysis, perception and
management, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 106:A9-A18.
Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling, SME Transactions,
316:43-50.
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
483 p (Oxford University Press: New York).
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model
for open pit mine planning: Application and risk analysis at a copper
deposit, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 116(3):A109-A118.
Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under
metal uncertainty – Application of stochastic mathematical
programming at an open pit copper mine and comparison to
conventional scheduling, in Advances in Orebody Modelling and
Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Ramazan, S and Dimitrakopoulos, R, 2004. Recent applications of
operations research in open pit mining, SME Transactions,
316:73-78.
Ramazan, S and Dimitrakopoulos, R, 2007a. Production scheduling with
uncertain supply: A new solution to the open pit mining problem,
COSMO research report, number 2, pp 257-294.
Ramazan, S and Dimitrakopoulos, R, 2007b. Stochastic optimisation of
long-term production scheduling for open pit mines with a new
integer programming formulation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 385-392 (The Australasian Institute of Mining and Metallurgy:
Melbourne)
Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
simulation, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 101:A104-A108.
Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse,
P, 2007. Blasor-blended iron ore mine planning optimisation at
Yandi, in Orebody Modelling and Strategic Mine Planning, second
edition (ed: R Dimitrakopoulos), pp 133-136 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Whittle, G, 2007. Global asset optimisation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 331-336 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Advances in Orebody Modelling and Strategic Mine Planning I
 HOME

Overall Optimisation of Yerilla Project

T Dincer1 and K Reynolds2

ABSTRACT offer significant technological and economical advantages in the

Yerilla Project, comprising the development of three relatively large
current project, including: increased leach feed grade after
nickel laterite deposits, is located 130 km northeast of Kalgoorlie. Boyce beneficiation of fines, increased range of ore types amenable to
Creek and Aubils deposits are located 25 km and 55 km respectively to leaching, lower acid and calcrete unit consumptions, lower nickel
the northwest of the project site, centred close to the main Jump-up Dam in circuit inventory, and full scale capacity could be attained with
deposit. Planning the development of multiple deposits added to the minimal ramp-up period.
challenges of a geologically and metallurgically complex project in the The key to the success of the Yerilla scoping study would be the
demanding industrial environment in early to mid 2008. The Yerilla integration of process performance and cost parameters into the
Project scoping study completed in September 2008 has resulted in an
mine optimisation process. The study completed in September
18 years’ production schedule, based on the processing of 3.2 million
tonnes resource per annum to produce average 21 000 tonnes of nickel
2008 (Heron, 2008b) has resulted in an 18 years’ production
and 1400 tonnes of cobalt per annum. The overall optimisation of the schedule based on processing of 3.2 million tonnes resource per
project has been crucial for making decisions during the study and is annum to produce average 21 000 tonnes of nickel and
based on the following interdependent technical studies: optimisation of 1400 tonnes of cobalt per annum. Conventional open pit mining
direct feed and beneficiation parameters, optimisation of cut-off grades method was selected for the development of the open pits based on
and throughput rates, mine sequencing and detailed production a major equipment feet of 180 t excavators and 90 t rear dump
scheduling, and analysis of mining equipment options, capital and trucks. Development of multiple deposits in parallel was required
operating costs. in the progress of the production schedule to supply a process feed
This paper presents the technical analysis approach taken in the Yerilla blend with consistent characteristics, which added to the
Project scoping study and the major findings. It has been demonstrated challenges of a geologically and metallurgically complex project.
that the balance between the interdependent project parameters, hence the The overall optimisation of the project was achieved to a
overall optimisation of the project, have been achieved at a sufficient sufficient level for a scoping study on the basis of the following
level through relatively basic analyses undertaken at the pit optimisation
interdependent technical studies: definition of beneficiation and
and mine sequencing stages of the project. The scheduling of equipment
capital and operating resources, depending on the economic environment atmospheric leach processing parameters, optimisation of cut-off
at the time of the study, would be of particular interest. and cut-over grades and throughput rates, mine sequencing of
multiple deposits and detailed production scheduling, and analysis
of mining equipment options, preliminary capital and operating
INTRODUCTION cost estimates.
Yerilla Project, comprising the development of three relatively
large nickel laterite deposits, is located 130 km northeast of
Kalgoorlie. Boyce Creek and Aubils deposits are located 25 km
and 55 km respectively to the northwest of the project site, centred
close to the main Jump-up Dam deposit (Figure 1). The project is
100 per cent owned by Heron Resources Limited (Heron),
involving the development of multiple open pit mines, a treatment
plant and associated utilities, services and infrastructure to recover
nickel and cobalt into a mixed sulfide product, using BHP Billiton
patented atmospheric leach technology (BHPB Patent
AU2007904228). Nickel-laterite mineralisation at the Jump-up
Dam, Boyce Creek and Aubils deposits is developed over
ultramafic dominated lithologies where basement rocks have been
deeply weathered to form a thick regolith profile, the character-
istics of which are summarised below in Figure 2. Most of the
high-grade nickel mineralisation is developed over serpentinised
ortho/ mesocumulate rocks in the form of goethite, smectite and
chlorite rich clays, while lower grade highly siliceous material is
common over adcumulate rocks. The thicknesses of the regolith
profile components and the mineralisation vary widely depending
on the underlying lithology, the presence of faulting and fracturing
in the bedrock, and the amount of surface stripping by erosion.
The initial Yerilla Project study completed in March 2008
(Heron, 2008a) concluded that while heap leaching was
technically feasible at that level of assessment, it carried high
technical risk and was not economically viable. The atmospheric
leaching in combination with beneficiation of fines was seen to

1. MAusIMM, Principal Mining Engineer, Mining Solutions

Consultancy Pty Ltd, 5 Salter Point Parade, Salter Point WA 6152.
Email: tdincer@miningsolutions.com.au
2. MAusIMM, Project Manager, Heron Resources Limited, Level 1, 37
Ord Street, West Perth WA 6005.
Email: kreynolds@heronresources.com.au FIG 1 - Yerilla Project location.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 159
T DINCER and K REYNOLDS
Regolith & Character
Alluvial Gravels Alluvial cover
Iron/Calcrete/ Calcrete/
Pisolitic Duricrust
Siliceous Cap Siliceous
Ferrugenous:
Smectite:
Ferrugenous/ varying ratios of
Nontronite and
Smectite Goethite, Haematite &
Montmorillonite Massive to
Kaolinite
Dissem.
Magnesite
& Dolomite
Varying ratios of Smectite,
Saprolite
Chlorite/Garnierite and Serpentine
Saprock Variabily oxidised Serpentine
Fresh Ultramafic Olivine/Serpentine
FIG 2 - Generalised regolith profile and mineralisation characteristics.
With the exceptions of Whittle Software used in pit limit
optimisations and MineSight, used as the host planning software,
all the other analysis and scheduling tasks in the study were
completed through advanced spreadsheets, custom programs and
a combination of standard procedures. Almost all the information
in the project has been linked electronically through procedures,
spreadsheets and basic data files, with references to the time and
team members involved in the project. The interdependent
project tasks as part of this approach were as follows: test work
program to obtain beneficiation performance data, analysis of the
beneficiation performance data and definition of processing
parameters, economic analysis of the processing methods and
resource utilisation, analysis of the cut-off grades and throughput
rates, mine production and processing schedules and mining
equipment selection, capital and operating cost estimates.
With the advancement of the computer technology, most of the
mine design and planning tasks, and consequently the decision-
making process in the mining projects, are largely dependent on
the specialised technical software chosen for the project. The
progress of the project commonly depends on the capabilities
and limitations of the specialised software and the experience of
the user of the software, especially in the production scheduling
area. Contrary to this, the sections below demonstrate how the
overall optimisation of the Yerilla project has been achieved
through a system of basic software tools that facilitated systemic,
relativity and sensitivity analyses. The Yerilla project was not
dependent on the capabilities of any specialised software at any
stage. The majority of the analyses for the optimisation of the
project parameters took place in the mine sequencing stage,
which is spatially and directly linked to the resource model,
mining areas and dependency between the mining areas (panels).
PROCESS PARAMETERS
Beneficiation and leaching flow sheet
Coarse low-grade gangue minerals (typically silica and magnesite
minerals) may be rejected from the ROM ore using relatively
simple scrubbing, attritioning and size separation techniques,
leading to an upgraded leach feed fraction. However, for the
Yerilla Project, not all ore types respond positively. In addition, as
head grade increases, increased nickel losses to the reject stream
can make beneficiation suboptimal. Those ore types not
160 Spectrum Series Volume 17
Over Print Ni % Co % FeO % MgO %
<0.5 <1.0 <0.05 >40 <20 0.5 to 5
0.3 to 0.5 to 0.05 to 25 to 10 to
0.5 to 35
Massive 1.5 2.5 0.5 60 35
to
Dissem.
Free
Silica 0.03 to
0.5 to 4 20 to 40 5 to 35
0.5
0.3 to 0.7 0.02 15 to 20 30
0.3 0.01 5 to15 30 to 45
beneficiated are processed via a direct leach grinding circuit ahead
of leaching. The atmospheric leach process assessed is a two stage
leach process, as patented by BHP Billiton (Patent: AU2007-
904228). In the first stage (primary) leach, ferruginous (or limon-
itic) ore types are leached under conditions of high acid addition,
in agitated tanks at atmospheric pressure and approximately
100°C. The residence time is relatively short at nominally three
hours. In the second stage (secondary) leach, saprolite ore types
are combined with the primary leach residue and leached for a
further 12 hours (nominal) at approximately 95°C under low acid
strengths. Under these conditions, dissolved iron from the primary
leach precipitates as jarosite, generating additional acid which is
utilised by the saprolite leaching reactions.
Beneficiation parameters
A program of test work was undertaken to define the beneficiation
performance of the Yerilla ore types. Beneficiation test work,
incorporating scrubbing and attritioning, has been undertaken on a
total of 34 samples. All major Jump-up Dam ore types and four
Boyce Creek samples were tested. As a result of the beneficiation
testwork programme, the selected beneficiation cut size and
performance parameters by ore type are presented in Table 1.
There are three main classifications: ferruginous, transitional, and
saprolite.
For the Yerilla Project, both ferruginous and transitional ore
types feed the primary leach circuit. A simplified flowsheet is
presented in Figure 3. Figure 4 summarises the beneficiation
performance results, including the average upgrade ratio verses
recovery performance for each ore type. The ore types with the
flat curves in the lower part of the graph are those which do not
show good beneficiation performance, being ferrougenous (F),
saprolite-nontronite/chlorite (S_NC) and saprolite-serpentinite
(S_S). These ore types were assigned for direct leach processing
without beneficiation. The ore types with the higher sloped
curves beneficiate well, including ferrougenous-nontronite/quartz
(F_NQ), Transitional samples (T), and saprolite-sepentinite/quartz
(S_SQ). The S_SQ sample set was found to have a bi-modal
distribution of good and poor performing samples so that the
average performance level was taken for this group.
Atmospheric leach parameters
For the optimisation work, leach performance was estimated
from preliminary data available from earlier studies and from
Advances in Orebody Modelling and Strategic Mine Planning I
 OVERALL OPTIMISATION OF YERILLA PROJECT

TABLE 1
Ore types and beneficiation performance.

Ore type Description Circuit Cut size Recovery (%)

μm
Mass Ni Co MgO Fe SiO2 AI2O3 CaO Mn
Ferruginous (containing goethite)
F Geothite + kaolinite +/- MnO Direct leach Grind 100 100 100 100 100 100 100 100 100
F_Q + Quartz Scrub & attrit 180 33.4 55.5 52.9 49.3 47.8 25.2 51.3 15.4 42.1
F_NQ + Nontronite or quartz Scrub & attrit 180 55.8 75.3 61.7 67.9 72.9 40.1 76.2 52.2 62.3
F3 + Carbonate Scrub & attrit 180 35.7 46.0 40.2 17.5 41.2 57.5 41.9 12.1 37.8
F_Q3 + Quartz + carbonate Scrub & attrit 180 33.4 55.5 52.9 49.3 47.8 25.2 51.3 15.4 42.1
F_NQ3 + Nontronite or quartz + carbonate Scrub & attrit 180 55.8 75.3 61.7 67.9 72.9 40.1 76.2 52.2 62.3
Transitional (containing geothite + serpentinite)
T_NCS + Chlorite + nontronite Scrub & attrit 180 39.0 61.6 62.6 41.1 58.2 37.3 63.0 47.9 48.8
T_NCSQ + Quartz + chlorite + nontronite Scrub & attrit 180 39.0 61.6 62.6 41.1 58.2 37.3 63.0 47.9 48.8
T_NCS3 + Chlorite + nontronite + carbonate Scrub & attrit 180 39.0 61.6 62.6 41.1 58.2 37.3 63.0 47.9 48.8
T_NCSQ3 + Quartz + chlorite + nont +carb Scrub & attrit 180 39.0 61.6 62.6 41.1 58.2 37.3 63.0 47.9 48.8
Saprolite (containing serpentinite)
S_S Serpentinite Direct leach Grind 100 100 100 100 100 100 100 100 100
S_S3 + Carbonate Scrub & attrit 500 66.8 75.8 81.4 70.6 78.9 56.2 72.0 71.0 81.6
S_SQ + Quartz Scrub & attrit 500 48.5 64.1 65.2 57.1 63.7 41.9 64.8 44.6 62.6
S_SQ3 + Quartz + carbonate Scrub & attrit 500 48.5 64.1 65.2 57.1 63.7 41.9 64.8 44.6 62.6
S_NC + Nontronite/chlorite Direct leach Grind 100 100 100 100 100 100 100 100 100
S_NC3 + Nontronite/chlorite + carbonate Scrub & attrit 500 83.2 91.8 93.7 83.0 87.4 81.2 85.6 48.3 87.7

Acid plant
Mining

ROM Stockpile
Crusher
Conveyor Benefication

Agitated Leaching Tanks

Crushed ore stockpile

waste
Fine ore after crushing

Limestone

Stage 1
Neutralisation
Recycling of process water

Precipitation
Stage
Filtering and drying

Acid addition

Intermediate Product
Bulk Bags

Zero discharge site

Shipping from Esperance

FIG 3 - Process flow sheet.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 161
T DINCER and K REYNOLDS

Average Ore Type Performance for Ni

Scrubbed and Attritioned Unit Operating Profit vs Feed Grade - S_SQ Ore Type
75, 180, 500, 1800, 3350 um cut size 100
1.80
90
80

1.60
70
60

Operating Profit ($/t ROM)

Upgrade Ratio

50
1.40 40 Beneficiation Feed
30 Direct Leach Feed
20
1.20 10
0
-10 0.0 0.2 0.4 0.6 0.8 1.0 1.2 1.4 1.6 1.8 2.0
1.00
-20
20.0 30.0 40.0 50.0 60.0 70.0 80.0 90.0 100.0
Distribution (%Ni) -30

F Scrub F_NQ Scrub T_all Scrub S_SQ good Scrub S_S Scrub -40
S_SQ bad Scrub S_NC Scrub F3 Scrub S-SQ3 Scrub F_NQ Attrit -50
F Attrit F3 Attrit T_all Attrit S_S Attrit S_NC Attrit
Feed Grade (%NiEq)
S_SQ good Attrit S_SQ bad Attrit S-SQ(3) all Scrub S_SQ(3) all Attrit S_SQ3 Attrit

FIG 4 - Beneficiation performance summary. FIG 5 - Cut-off and cut-over grade determination.

knowledge of typical performance from other projects. Leach
extraction parameters used in the optimisation work are
presented in Table 2. Final leach test work data was not available
until late in the study schedule.
TABLE 2
Leach performance assumptions by ore type.
Ore type Circuit
Ferruginous Primary
Transitional Primary
S_S and S_S3 Secondary
S_SQ and S_SQ2 Secondary
S_NC and S_NC3 Secondary
CUT-OFF GRADE AND THROUGHPUT RATE
Break-even cut-off and cut-over grades
Subsequent to the determination of the metallurgical parameters
for the ore types, economical comparison of the beneficiation plus
leach and direct leach processing methods was necessary for
further classification of the ore types for mine planning purposes.
This required the calculation of the cut-off and cut-over grades
depending on the operating costs and recoveries for each process-
ing method. The cobalt credit for the ore types was included in the
calculations. Leach plant capital reduction benefit for the
beneficiation process was included in the comparison of the
economic values for the calculation of the cut-over grades between
the beneficiation plus leach and direct leach circuits. The capital
reduction benefit was based on the fact that a lower leach plant
capacity would be required to produce the same amount of metal
as a result of the upgrade through the beneficiation process. The
effect of beneficiation is estimated as a 12.5 per cent overall
reduction in the processing plant and related infrastructure capital
costs. Figure 5 shows the variation of the unit operating profit as
the feed grade varies for two processing methods. The curves are
for the S_SQ ore type with the most significant tonnage (22 Mt).
The graph shows that the beneficiation adds value relative to direct
leach option for the low-grade ore below ~1.00 per cent Ni by
upgrading the metal content. Above 1.00 per cent Ni feed grade,
direct leaching of the ROM ore is better since the metal losses
become more prominent and start reducing the unit operating
value compared to the direct leach option. Based on Figure 5, the
cut-off and cut-over grades for the S_SQ ore type are as follows:
0.55 per cent Ni break-even cut-off grade for the beneficiation
ROM feed, and 1.10 per centNi cut-over grade for the direct
leach ROM feed.
The same methodology was applied to determine cut-off and
cut-over grades for the other ore types stored in the resource
model. The break-even cut-off grades generally varied between
Recovery (%) 0.40 per cent Ni and 0.70 per cent Ni, depending on the process
Ni Co recoveries for the ore types. In order to simplify ore classification
95 90
criteria, the cut-off and cut-over grades were standardised across
the ore types after checking the economic sensitivities.
95 90
75 60 Cut-off grade analysis
85 70
Prior to the determination of the ultimate pit limits and
85 70 production scheduling, a cut-off grade and throughput optimis-
ation study was completed as part of the pit limit optimisations to
determine the optimal production parameters for the project. The
four cut-off grade cases in the study included the break-even
(BE), 0.6 per cent Ni, 0.7 per cent Ni and 0.8 per cent Ni cut-off
grades. Preliminary production schedules were completed for each
cut-off grade case based on the optimal pit shell sequences
combined from Jump-up Dam and Boyce Creek models available
at the time of the analysis. The combinations of 2.0 Mt/a, 2.5 Mt/a
and 3.0 Mt/a leach plant throughput rates were studied for each
cut-off grade case.
The initial schedules accommodated the variable cut-off grades
through time as the cut-off grades for each increment in the
mining sequence was selected to provide maximum Net Present
Value (NPV). However, the project NPV was not significantly
better than the constant cut-off grade cases with relatively higher
NPVs. In fact, the overall differences between the studied constant
cut-off grade cases were not so large (only a few percentage
points) due to the beneficiation of the low-grade ore within a
limited grade range below the cut-over grade. Also considering the
practicality aspects of the project associated with the number of
ore types and relatively lower confidence in the early stage of
resource definition, especially for Boyce Creek and Aubils
deposits, a constant cut-off grade strategy was opted in the
production schedules.
The graph in Figure 6 shows the variation of the cumulative
operating NPV at Years 4, 8 and 12 of the preliminary production
schedules for 2.5 Mt/a throughput rate. The analysis of the cut-off
grades across the throughput rates can be summarised as follows:
Compared to the break-even cut-off, the NPV improvement in the
0.6 per cent Ni cut-off cases is generally not significant, indicating
that some marginally economic resource is still included in the

162 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 OVERALL OPTIMISATION OF YERILLA PROJECT

2500ktpa Leach Cumulative Operating NPV improvements in the operating NPV as the throughput rate
140% increases, the capital cost of the processing plant capacity
0.8%Ni COG
increase needs to be taken into account. The same curves in
135%
0.7%Ni COG Figure 8 are based on the NPV values normalised for the plant
130% 0.6%Ni COG and infrastructure incremental capital costs. The throughput rate
B.E. COG
125% analysis for the normalised NPVs showed that the overall NPV
Operating NPV (%)

120%
of the project would not be improved significantly above a
2.5 Mt/a throughput rate. The NPV for the 3.0 Mt/a throughput
115%
rate is significantly better at Year 8, but drops to a lower level
110% than the 2.5 Mt/a case at Year 12 of the schedule. The shorter
105% mine life for 3 Mt/a case would also affect the depreciation of the
100%
plant and infrastructure capital, slightly increasing the overall
(Year 12) NPV gap between 2.5 Mt/a and 3.0 Mt/a cases. Subject
95%
to further review in future, the throughput analysis results
90% showed that a 2.5 Mt/a throughput rate would satisfy the
3 4 5 6 7 8 9 10 11 12 13
production requirements for the project and improve the project
Years
NPV slightly in the first 12 years of operations.

FIG 6 - Operating NPV for cut-off grade cases. 0.7%Ni Cut-off Normalised Cumulative Operating NPV
140%
3000ktpa Leach
schedules; Compared to the 0.6 per cent Ni cut-off, the 0.7 per 135% 2500ktpa Leach
2000ktpa Leach
cent Ni cut-off grade provides significant improvement in the 130%
operat- ing NPV values, eight per cent in Year 8 and four per cent
125%
in Year 12; and Although the 0.8 per cent Ni cut-off grade

Operating NPV (%)

improves the NPV more significantly at the end of Years 4 and 8 120%

(12 per cent and 16 per cent respectively), the overall NPV is 115%
lower than the 0.7 per cent Ni cut-off case due to the
110%
under-utilisation of the available resources. The shorter mine life
would also reduce the depreciation time for the plant and other 105%

project infrastructure. 100%

The low-grade resources below cut-off grade were stockpiled 95%

but not considered for treatment within the scope of the project
90%
due to the marginal economics, long mine life (close to 20 years) 3 4 5 6 7 8 9 10 11 12 13
and the insignificant impact on the project NPV. The processing Years
of the low-grade resources would require the feasibility of a new
follow-up project with the increased beneficiation and leaching FIG 8 - Normalised NPV for throughput rate cases.
capacity to achieve sufficient levels of metal production rates.
Further potential resource definition in the project area, as well
as likely changes in the mineral economics in the 20+ years time Cut-off grade and throughput optimisation
frame, did not justify the inclusion of the low-grade treatment results
within the scope of the current project.
The cut-off grade and throughput rates analyses discussed in the
Throughput rate analysis two subsections above are related and all the combinations of the
cut-off and throughput cases were studied. The following is a
For the selected 0.7 per cent Ni cut-off grade, the variation of the summary of the combined findings from the review of cut-off
cumulative operating NPV at the end of the Years 4, 8 and 12 of grade and throughput rate production analysis results:
the production schedules can be seen in Figure 7 for the leach
capacity cases studied. Although the curves show significant • the proposed 0.7 per cent Ni cut-off grade removes the
marginally economical resources from process feed and
brings some cash flow earlier without compromising the
0.7%Ni Cut-off Cumulative Operating NPV overall value of the project;
140% • the cut-off grades above 0.7 per cent Ni reduce the proportion
3000ktpa Leach
135% 2500ktpa Leach
of the lower grade beneficiation feed significantly, hence
2000ktpa Leach reducing the overall utilisation of the low-grade resources;
130%

125%
• the beneficiation of the low-grade ore does not improve the
operating NPV of the project, but helps to reduce the leach
Operating NPV (%)

120%
plant capacity (hence lower capital) to achieve the same
115% nickel production;
110% • the review of the resource models have shown that the
105% resource outlines become less continuous as the cut-off grade
100%
increases above 0.6 per cent Ni, which would in turn increase
the low-grade dilution, or require more selective mining; and
95%

90%
• the 2.5 Mt/a total leach rate case produces higher, or at least
3 4 5 6 7 8 9 10 11 12 13 similar overall NPV (including capital), compared to the
Years 2.0 Mt/a and 3.0 Mt/a cases.
As a project decision based on the findings above, the
FIG 7 - Operating NPV for throughput rate cases. combination of the 0.70 per cent Ni cut-off grade and 2500 kt/a

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 163
T DINCER and K REYNOLDS

total leach feed rate provides the best value for the project. The
beneficiation rate, secondary/primary leach rates and annual
nickel production for this combination, calculated as part of the Jump-up Dam
analyses, are as follows: 1.4 Mt/a of average beneficiation rate
providing 0.7 Mt/a of beneficiated leach feed rate, 1.8 Mt/a of
average direct leach feed rate for a total of 2500 kt/a total leach
rate, and 22 000 t/a average nickel production in the first
13 years, then 20 000 tpa average for the rest of the mine life.
Note that the Aubils resource model was not ready and not
included in the study at the time of the production analysis. It is
currently planned to develop the Aubils deposit after eight years
of plant operation, so the results from the production analysis
above will be critical for cash flow in the early years. Reviews of
the production schedule and Aubils open pit economics indicated
that the inclusion of the Aubils deposit is likely to strengthen the
case for the selected cut-off grade and throughput rate.

PRODUCTION SCHEDULE
Mine sequencing and scheduling
Detailed mine designs were not justified at this stage of the study
due to the generally shallow nature (40 - 80 m depth) of the pit
shells. The pit slopes allowed for the ramps and the smoothing
effect in the pit design process would not be significant on the
reported quantities, considering the relatively large extend of the
resource areas. Although the Aubils pit quantities, due to the
discontinuous nature of the open pit limits, may be affected
significantly as a result the pit design process, this deposit was
scheduled late in the mine life and the resource definition was in
an early stage at the time of the study. The optimal pit areas for
each deposit have been divided into the production panels (as
development stages) that can be developed relatively indepen-
dently. The internal variation in the optimal pit shell surfaces
(highs and lows) were taken into account in the definition of the
panel boundaries to form a practical mining sequence. The
economically more significant (low cost or higher value) areas
were delineated within the panel boundaries, depending on the pit
optimisation results. The resultant production panels for the
Jump-up Dam, Boyce Creek and Aubils deposits can be seen in
Figure 9. Each panel represents three to six months of mine
production in terms of tonnages, but can be developed within six
to 12 months due to the mining of multiple panels for blending
purposes and limitations with the vertical advance rates.
The production scheduling process allowed mining of the
multiple panels from multiple deposits to achieve the schedule
objectives. An initial mining sequence was defined at the start of
the scheduling process based on ranking the economics of the
production panels. The initial sequence providing the best
operating NPV was modified in the scheduling process,
depending on the priority of the objectives and constraints.
Consistent primary and secondary leach feed ratios, consistent
utilisation of the beneficiation plant capacity, variation of grades
(Ni, Co, CaO, Mn, FeO, SiO2, MgO, CO3, Al2O3) within a
reasonable range, deferment of the mining costs and earlier higher
cash flow were the main objectives of the production schedule.
The production scheduling methodology comprised four phases in
order to balance the mine, mill and stockpiling schedules as
follows: an initial ore supply schedule was completed to balance
the production of the ore types to achieve relatively smooth
primary to secondary leach ratios around 0.9, the processing
schedule was adjusted to achieve relatively constant beneficiation
to direct leach ROM feed ratios through project life, the
processing and stockpiling schedules were finalised within the
grade range constraints in the third phase to achieve relatively
smooth (decreasing in stages) nickel production over mine life,
and in the final phase, the production from different panels and
deposits was balanced to regularise the mining rates with the
working stripping ratio increasing gradually over mine life.
The details provided in the production schedule are considered
sufficient to highlight any potential operational difficulties
Boyce Creek Aubils
FIG 9 - Operating NPV for cut-off grade cases.
throughout the mine life, as well as to provide a basis for the
long-term cash flow analysis. The resultant long-term production
schedule is relatively flexible, and the short-medium term
production tonnage and grade variations around the long-term
schedule are expected to lie within an acceptable range.
Production schedule results
The total tonnes and ROM ore tonnes scheduled by deposits can
be seen in Figure 10. The annual production rate increases
gradually in the schedule as the areas with higher stripping ratios
developed towards the end of the mine life. The variation in
mining rates was smoothed, but is still largely dependent on the
processing requirements rather then the maximum utilisation of
the mining equipment (discussed in mining equipment section
below). The ROM ore production in the schedule varies between
3.0 Mt/a and 3.5 Mt/a, depending on the beneficiation feed rates,
mining rates and variation in the stockpile balances. Figure 11
shows that the beneficiation rates vary between 1.2 - 1.7 Mt/a
and the direct leach feed rates vary between 1.7 - 2.0 Mt/a. The
beneficiation feed ROM grade averages at 0.81 per cent Ni
through mine life and the average grade for direct leach ROM
ore is 1.05 per cent Ni. The beneficiated leach feed grade is
slightly higher than the direct leach feed grade after 30 - 35 per
cent upgrade in the beneficiation process. Beneficiated leach
feed constitute 27 per cent of the total leach feed in the schedule,
providing 20 per cent reduction in the leach plant capacity and
related operating cost savings in low-grade ore processing.
As seen in the last graph in Figure 11, the schedule achieves
relatively constant primary and secondary leach feed tonnages
and grades throughout mine life. The primary and secondary
leach feed grades will be similar in the first seven years when the
relatively higher-grade areas of the Jump-up Dam and Boyce
Creek deposits will be mined. The secondary leach feed grades

164 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 OVERALL OPTIMISATION OF YERILLA PROJECT

Total Tonnes Mined by Deposit Beneficiation and Direct Leach ROM Ore Feed
4,000 1.60
22,500
20,000 3,500 1.40

17,500 3,000 1.20

G rade (N i% )
Tonnes'000
15,000 2,500 1.00
Tonnes’ 000

12,500 2,000 0.80

10,000 1,500 0.60
7,500
1,000 0.40
5,000
500 0.20
2,500
0 0.00
0
-1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
-1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Years
Years
Direct Leach ROM Feed Beneficiation Feed ROM Feed
Jump-up Dam Boyce Creek Aubils Direct Leach ROM Grade Beneficiation ROM Grade

Beneficiated Leach and Direct Leach Feed

ROM Ore Tonnes Mined by Deposit 3,000 1.50
4,000
2,500 1.25
3,500
3,000 2,000 1.00

Grade (Ni% )
Tonnes'000
2,500
Tonnes’ 000

1,500 0.75
2,000
1,000 0.50
1,500
1,000 500 0.25
500
0 0.00
0 -1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
-1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 Years
Years Direct Leach ROM Feed Beneficiated Leach Feed
Direct Leach ROM Grade Beneficiated Leach Feed Grade
Jump-up Dam Boyce Creek Aubils
Primary and Secondary Leach Feed
3,000 1.50
FIG 10 - Mine production schedule.
2,500 1.25

2,000 1.00

G rade (N i% )
Tonnes'000

will be slightly lower after Year 7, when the remaining resource

1,500 0.75
with relatively low-grade high saprolite proportion is blended in
to the Aubils resource with high primary feed proportion. The 1,000 0.50

annual nickel production in the schedule generally varies 500 0.25

between 22 000 and 24 000 tonnes (Figure 12). The average
0 0.00
cobalt production is 1400 tonnes per annum, raising the total -1 1 2 3 4 5 6 7 8
9 10 11 12 13 14 15 16 17 18
metal production close to 25 000 tonnes per annum except the Years
last four years. The total metal production gradually drops to Primary Leach Feed
Primary Leach Feed Grade
Secondary Leach Feed
Secondary Leach Feed Grade
21 000 tonnes of metal in the last four years when the low-grade
and high stripping ratio areas of the three deposits will be mined.
Note that these metal production values are based on the FIG 11 - Beneficiation and leach feed schedules.
preliminary leach recovery estimates, which have been reduced
by approximately ten per cent in the final estimates used in the Annual Nickel and Cobalt Production
project’s financial analysis. 30.0

25.0
MINE EQUIPMENT AND COSTS ANALYSIS
20.0
Tonnes’ 000

The major mining equipment selected for the project comprises

15.0
180 tonne excavators and 90 tonne capacity rear dump trucks. The
excavator dipper capacity was selected ten per cent larger than the 10.0

standard capacity, depending on the lighter density of the laterite 5.0

and other oxide material. The variation of the excavator and truck
0.0
numbers through mine life are shown in Figure 12. The mining -1 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18
fleet numbers increase following mainly the increase in the Years

stripping ratio and partly in the number of deposits mined
simultaneously. Considering the number of deposits and panels to
be mined in parallel, the equipment fleet has been standardised
despite the uneven increase in the equipment requirements
throughout mine life. One of the excavators was scheduled for a
single shift per day in some years and the proposed equipment
numbers allow forward purchase of the equipment to provide
additional flexibility in the production rates.
As shown in Figure 12, the excavator and truck fleet utilisation
in the schedule provides 20 - 30 per cent extra capacity through
mine life for flexibility and effective on-the-job utilisation. The
waste and/or ore mining campaigns will/may also demand
additional equipment capacity up to several months during
operations. Although the available equipment capacity should
reduce the stockpiling requirements, the construction and
Nickel Ni Eq. Cobalt
FIG 12 - Nickel and cobalt production schedule.
reclamation from the relatively large stockpiles may be necessary
to regulate the processing feed quantity and qualities (especially at
Boyce Creek and Aubils, where the ore needs to be rehandled
anyway). Contrary to the approach discussed above, in a tight
equipment capacity case, the nature of the laterite operations is
likely to require the mobilisation and hire of additional equip-
ment to overcome production difficulties. A preliminary cost
comparison indicated that the capital cost for the additional
equipment capacity would be justified especially for effective (on
the job) utilisation of the equipment, saving fuel and labour costs

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 165
T DINCER and K REYNOLDS
No of Excavators, Left Axis
No of Trucks, Left Axis
20
Scheduled Time Utilisation, Right Axis
18
Available Time Utilisation, Right Axis
16
14
No of equipment
12
10
8 40%
6 30%
4 20%
2 10%
0 0%
-1 1 2 3 4 5 6 7 8 9 10 11 12 13
Years
FIG 13 - Mining equipment requirements.
through mine life. Reducing the non-productive operating time of
the mining equipment would also have the benefit of reducing the
ancillary equipment utilisation and areas required for large
stockpiles. The NPV of the capital cost of purchasing mining
equipment earlier in the schedule and under-utilisation was
compared to the NPV of the potential rehandling costs. This
analysis indicated that the NPV of the additional equipment
capacity at all times is approximately equal to the NPV of the
rehandling of 20 per cent of the ROM ore. This would drop to less
than ten per cent rehandling of the ROM ore when the additional
capital cost of the stockpile rehandling equipment (eg an extra
loader and two trucks) was considered in the analysis.
CONCLUSIONS
The economic analysis of the Yerilla Project indicated that an
increase in the capital costs for a reduction in the operating costs
would improve the value of the project – simultaneously resulting
in the planning of more environmentally responsible mining
operations. This result is not surprising considering the economic
conditions and trends in other industries at the time of the study.
The scoping study results do not claim to solve all the issues,
especially related to the variability in the project. The results are
mostly limited with the available information, which needs to be
reviewed, revised and further detailed as the confidence on the
resources and other project parameters increase.
The approach taken and the results obtained in the Yerilla
project have demonstrated that overall balance between the
geological, metallurgical and mining parameters in a project can
be achieved through basic analyses linked spatially to the resource
and mining areas. The overall optimisation of the Yerilla Project
was effectively achieved in the scoping study through:
understanding the spatial geological and metallurgical variability,
systematic analysis of the cases/options, task or problem
isolation, relative analyses and comparisons, sensitivity analysis
on the findings to further test the variability, and concluding on the
trends and ranges with known limitations rather than specific
numbers.
166
Although the approach above is sometimes called
‘conventional’, ‘old fashioned’ or ‘manual’ by advocates of
100%
specialised mining software, the system of procedures, data and
90%
analysis tools can be linked and executed to reach a project
80%
balance efficiently. Contrary to this approach, it is common that
70%
the planning of the laterite (and other complex) deposits is
Utilisation (%)
60%
generally limited with the capabilities of the available
50%
optimisation/scheduling software and experience of their users.
The first objective of project development, especially at the
scoping stage, should be to achieve an overall balance between the
project parameters. Such a project balance, with associated
limitations, would also provide scope for the employment of any
14 15 16 17 18
specialised software for further project development. Considering
the usual limitations with the available data at the project stage,
and future economic uncertainties through the life of the project,
being ‘approximately right practically’ in an overall project sense
should be much better compared to being ‘absolutely right
theoretically’ in a specific area of the project. It is also suggested
that results from optimisation studies, especially production
schedule related studies, should be compared to a practical base
case to acknowledge the contribution of the software and justify
any changes, if any, for the increased complexity. The Yerilla
Project experience and the results obtained have also again shown
(after Dincer and Peters, 2001; and Dincer, 2001) the importance
of the basic pit limit optimisation and mine sequencing processes
in mine planning. Considering the interdependency of the project
parameters and iterative analyses required to reach a balance,
these two basic processes can effectively: link geological,
metallurgical, mining and economical parameters both
structurally and procedurally; provide a range of physical and
economical quantities with spatial/geometrical references for the
whole resource area; facilitate a wide range of evaluations and
sensitivity analyses on the project parameters; and potentially
provide foundations for an optimised mining project and
production schedule.
ACKNOWLEDGEMENTS
The authors would like to thank Heron Resource Limited for
permission to publish this paper. The authors would also like to
acknowledge Mr James Ridley for his contribution to the project
in the areas of geology and resources.
REFERENCES
BHP Billiton, 2007. Atmospheric acid leach process for laterites
Australian patent: AU2007904228.
Dincer, T, 2001. Application of pit optimisation algorithms beyond open
pit limits, in Proceedings 17th International Mining Congress and
Exhibition of Turkey, pp 549-556.
Dincer, T and Peters B, 2001. Blending optimum pit mining sequences,
in Proceedings Strategic Planning Conference, pp 43 - 54 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Heron Resources Ltd, March 2008. Jump-up Dam project prefeasibility
study, internal report.
Heron Resources Ltd, September 2008. Yerilla project scoping study,
internal report.
Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I

Investigating the Variability of Key Coal Quality Parameters in
Continuous Mining Operations when Using Stockpiles
J Benndorf1
ABSTRACT
The delivery of in-spec coal qualities is essential for an efficient and
environmental friendly operation of modern coal-fired power plants. The
design of the mining operation and blending opportunities plays a key
role in homogenising variability and improving the prediction of key
quality parameters, such as the calorific value (CV). The presented case
study uses the technique of conditional simulation in geostatistics to
investigate the variability of the CV and its behaviour along the
extraction, transportation and blending process in a continuous mining
environment. Results of the study enable the opportunity to optimise the
operation mode and adjust it according to the customer’s requirements.
INTRODUCTION
The efficient and environmental friendly operation of modern
power plant technologies calls for strictly meeting the customer’s
specification of key coal quality parameters while maintaining a
low variability of these parameters. The daily mine production
aims to meet the customer specifications in terms of tonnage and
coal quality parameters for different products while minimising
specific costs. The key objective of complex continuous mining
systems is to achieve in-spec deliveries at the lowest cost, by
optimal blending from different faces and multiple seams with a
diverse quality distribution. The blending process typically
involves the operational mode of the excavators, the conveying
system, as well as stacking and reclaiming at stock and blending
yards. The application of conditional simulation in coal mining
operations is not very well documented, with the exception of
some simplified case studies (Hohn and McDowel, 2001; Costa,
Zingano and Koppe, 2000; others). The technique of conditional
simulation (Goovaerts, 1997) with its ability to map in-place
variability and uncertainty in the prediction of coal quality
parameters is suited to investigate the ability of the
homogenisation effect of mining systems. Case studies for grade
control, stockpiling and blending strategy optimisation in diffuse
deposits demonstrate the benefits the application of conditional
simulation can generate (eg Benndorf and Dimitrakopoulos, 2010,
in this volume).
This contribution investigates, based on conditional simulation,
the ability to homogenise coal quality parameters in the process of
digging, conveying and stocking in a continuous mining
environment. Objects of study are, in detail, the different excavator
types (bucketwheel excavator and chain excavator) as well as the
mode of operation of a strata type stockpile. A case study is
presented in a lignite mine in Eastern Europe and conclusions
follow.
DESCRIPTION OF THE CONTINUOUS MINING
SYSTEM UNDER STUDY
Figure 1 illustrates the continuous mining system considered in
this case study. The system contains two excavators, a chain and
a bucket wheel, which are positioned in different cuts at separate
belt conveyors. Both conveyors combine to join a central
stationary conveyor before the coal is loaded onto trains. There is
also the option to stack the coal in a coal stock and blending yard
with the goal being to homogenise the variability of coal quality
parameters before shipping it.
1. GAusIMM, MIBRAG MBH, Germany.
Email: JoergBenndorf@gmx.de
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
1. cut
2. cut
Coal stock and
blnding yard
Train load
FIG 1 - The continuous mining system under study.
In total, the mining system can be operated in four operating
modes:
1. operating only the first cut with a bucket wheel excavator
without utilising the blending yard,
2. operating only the second cut with a chain excavator
without utilising the blending yard,
3. operating both cuts at the same time without utilising the
blending yard, and
4. operating both cuts at the same time and utilising the
blending yard.
The study investigates the accuracy or certainty in prediction
and the variability of the key quality parameter calorific value
(CV) of the shipped coal as a function of the operating mode.
The study is based on 25 realisations generated by conditional
simulation of the spatial distribution of the CV. A mining unit of
1500 t is considered, which reflects a train load.
Figure 2 compares the digging mode of the two excavator
types (chain and bucket wheel). Typically the bucket wheel
excavator is positioned on a bench and digs the coal above the
bench in a high cut mode in horizontal slices of about 5 m. This
enables a selective extraction of the typically horizontally
deposited coal strata. The selectivity is limited by a minimum
mining thickness. In a deep cut mode operation, the chain
excavator digs the face in vertical slices along the whole cutting
depths. This causes a blending effect of the horizontally
deposited coal strata while digging, and limits its ability to
selectively mine this strata. Figure 3 shows photos of a bucket
wheel and chain excavator.
To analyse the variability of the coal quality parameters, block
models were generated, which reflect the digging mode of both
excavators.
ANALYSIS OF THE ABILITY TO CONTROL COAL
QUALITY PARAMETERS WITHOUT UTILISING
THE BLENDING YARD
This section focuses on the described operation modes 1 to 3. It
is assumed that the mining equipment is successively digging the
Spectrum Series Volume 17 167
J BENNDORF

Digging mode of different excavator types

Chain excavator
Bucket wheel excavator

FIG 2 - Digging modes of the two considered excavator types (bucket wheel and chain excavators).

FIG 3 - Photos of (left) a bucketwheel excavator and (right) a chain excavator (MIBRAG, 2009).

Prediction of the CV value of coal extracted by a bucket-wheel-excavator Prediction of the CV value of coal extracted by a chain-excavator
14,0 16,0

12,0 14,0
Calorific value (CV) in MJ/kg

Calorific value (CV) in MJ/kg

10,0 12,0

8,0 10,0

6,0 8,0

4,0 6,0
1.000 6.000 11.000 16.000 21.000 26.000 31.000 36.000 1.000 6.000 11.000 16.000 21.000 26.000 31.000 36.000
Cumulated extracted tonnage Cumulated exctracted tonnage

FIG 4 - Predicted alorific value (CV) of coal delivered from the first FIG 5 - Predicted CV of coal delivered from the second cut utilising
cut utilising a bucket wheel excavator. a chain excavator.

blocks in the digging mode explained in the previous section at a operating mode 2, that is the chain excavator, in the second cut.
constant digging rate. Using the generated block models, the CV The figures show the prediction based on 25 realisations (thin
of the extracted coal is predicted. Figure 4 shows the predicted gray lines), the average of the realisations (thick line) as well as
CV for operating mode 1 that is the bucketwheel excavator in the the accuracy (certainty) in prediction by the means of the 90 per
first cut. Figure 5 shows respectively the predicted CV for cent confidence level (dashed lines).

168 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 INVESTIGATING THE VARIABILITY OF KEY COAL QUALITY PARAMETERS IN CONTINUOUS MINING OPERATIONS

The comparison of both figures illustrates the abilities and Vorhersage des Heizwertes
Prediction of the CV of(roh)
coalbeim
whengemeinsamen Betrieb
operating both beider Geräte
excavarors

limits of both types of equipment in selectively extracting and 14,00

homogenising the in-place variability of coal quality parameters. 13,00

When operating the bucket wheel excavator, a cyclic behaviour of
12,00
the CV can be detected. The reason is the combination of a

MJ/kg
strongly vertically varying CV and the vertical digging of slices

MJ/kg
11,00

in in
per digging location (set up). Due to the strong horizontal

value (CV)
10,00

CalorificHeizwert
correlation of the CV, a very similar cycle is repeated in the next 9,00
set up of the equipment. When operating a chain excavator, the
8,00
strong varying vertical CV values are homogenised over the whole
digging depth. The resulting flow of extracted coal is characterised 7,00

by a low variation of CV. 6,00

1500 6000 10500 15000 19500 24000 28500 33000 37500
Two criteria are used herein to evaluate the homogenisation Kumulativ
Cumulatedgeförderte Tonnage
extracted tonnage

effect in the different operating modes: the accuracy or certainty in

the prediction of a unit of 1500 t (train load), measured by the
standard deviation of the 25 realisations, and the difference
between two successive extracted units as an expression of
variability. The difference in CV of successive delivered units is
typically a criterion, which is part of the customer’s specification.
Table 1 documents the results in terms of the criteria for the two
considered cases. The differences in the uncertainty of predicting
the CV between the two equipment types are due to the digging
mode, but also due to different areas of the mined deposit with
different structural behaviour. However, the bucket wheel
excavator shows significantly higher values of the uncertainty in
predicting the CV compared to the chain excavator. The reason is
the occurrence of local blocks, which are due to the exploration
level less certain in the prediction of CV. Due to the selective
extraction with the bucket wheel excavator, these blocks cannot be
blended with more certain blocks, as is possible with the chain
excavator. Considering the second criteria, the difference between
two successive extracted units, the selectivity of the bucket wheel
excavator is clearly illustrated. With an average of approximately
0.72 MJ/kg between the 1500 t units, the difference is about 80 per
cent higher compared to the chain excavator. The maximum
difference can be up to 7.95 MJ/kg.
Figure 6 shows the prediction of the CV when operating both
excavators at the same time (operating mode 3). Thereby with
3000 m3/h for the bucket wheel excavator and 1500 m3/h for the
chain excavator, the two different digging rates are taken into
account. The results in terms of the two criteria, uncertainty in
prediction of a unit of 1500 t (train load) and the difference
between two successive extracted units, are shown in Table 1. The
blending of the two material flows from both excavators shows
two effects. The uncertainty in predicting the CV of the shipped
coal decreases compared to only operating one of the two
excavators. On average, the standard deviation decreases by six
per cent compared to the chain, and 16 per cent compared to the
bucket wheel excavator. The second effect is a reduction in
variability between two successive units compared to the bucket-
FIG 6 - Predicted CV of coal delivered from both excavators
operated at the same time.
wheel excavator. On the other hand, this difference increases when
compared to only operating the chain excavator.
To generalise, an operation of multiple excavators at the same
time decreases uncertainty in predicting the coal quality
parameters of shipped units. At the same time, extreme
variability is smoothed. This homogenisation reduces differences
between two successive extracted units. With the goal of
selectively extracting coal with a certain specification, the bucket
wheel excavator should be operated separately.
ANALYSIS OF THE ABILITY TO CONTROL COAL
QUALITY PARAMETERS UTILISING THE
BLENDING YARD
The key function of a coal stock and blending yard is, besides the
buffering, the homogenisation of coal quality parameters. This
transforms the characteristics of the incoming material flow in an
outgoing material flow whose characteristics are defined by
customer specifications and may be of contractual relevance
(Figure 7). The characteristics of the incoming material flow are
a function of the geological conditions, the applied selectivity in
extracting the deposit, and the operation mode in the pit.
The following considerations concern the homogenisation
effect. The efficiency of blending and smoothing variability is
significantly dependent on constructive parameters as well as the
operation of the blending yard. Constructive factors are hereby the
type of the yard and its length and width. The following
constructive parameters are assumed in the case study: the
blending yard is of type ‘strata’ (Figure 8). The coal is stacked
layer by layer in six sub yards, which are each 120 m in length and
50 m in width. Figure 9 illustrates the mode of operation of a
‘strata’ blending yard. The coal coming from the pit (incoming

TABLE 1
Summary of the criteria for measuring uncertainty in prediction of a unit of 1500 t (train load), and the difference between two
successive extracted units for different operating modes.

Information in this table are of unit MJ/kg Bucket wheel excavator Chain excavator Operating both
(1. cut) (2. cut) excavators
Certainty of predicting Average standard deviation 0.71 0.66 0.62
the CV value of 1.500 t
Min standard deviation 0.26 0.27 0.41
Max standard deviation 1.67 1.05 1.00
Average 90 per cent confidence level 1.87 1.92 1.76
Min 90 per cent confidence level 0.74 0.83 1.29
Max 90 per cent confidence level 4.41 2.99 2.90
Differences in CV Average differences between two successive 0.72 0.40 0.60
between successive units
extracted units
Min difference between two successive units 0.00 0.00 0.00
Max difference between two successive units 7.95 1.84 3.30

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 169
J BENNDORF
in-coming
material flow Blending yard
(from pit)
FIG 7 - Process of blending in a coal stock and blending yard.
Stacker
FIG 8 - Coal stock and blending yard of type strata (MIBRAG, 2009).
material flow) is stacked into layers, which are spread along the
total length of a subyard by a continuously up and down moving
stacker. The number and thickness of the several layers is variable
and can be influenced by the moving rate of the stacker dependent
on the total production rate of the mine. The yard is reclaimed in a
fan-shaped manner orthogonally to the alignment of the stacked
layers by a scraper. In this way the coal quality of the outgoing
material flow is formed as an average over the total number of
stacked layers.
The number of total layers in the yard can be varied by the
mode of operation. The effectiveness of homogenisation shall be
investigated subsequently as a function of the number of layers.
Two criteria are used to evaluate the results:
1. the effect of homogenisation, which is defined herein as the
ratio between the dispersion variance of the CV of the
incoming material flow sin and the dispersion variance of the
CV of the out coming material flow sout (Equation 1) and;
2. the expected improvement of certainty in predicting the CV
of a unit of 1500 t, which is defined herein as the ratio
between the average standard deviation in predicting units of
1500 t before stacking, and the average standard deviation in
predicting units of 1500 t after reclaiming (Equation 2).
S in
effect of homogenisation = (1)
S out
σ 2 out
improvement of certainty in prediction = (2)
σ 2 in
170 Spectrum Series Volume 17
out-going
material flow
(delivery)
Reclaimer
Description of the blending yard model and
assumptions
The model is based on the prediction of the CV of the extracted
coal while operating both excavators (operation mode 3) using
25 simulated realisations of the block model. Every single
realisation can be seen as a possible scenario of the coal quality
parameter CV of the produced material flow coming out of the
pit and being stacked into the blending yard. Figure 10 illustrates
the model of the blending yard. The idea is based on a study
shown in Benndorf (2009), and Pavloudavis and Agioutandis
(2001). The model is composed of n columns and m rows. The
number of layers in the yard is m. The single pockets are of size
of 300 t and represent a fraction of the considered units of 1500 t.
Based on the assumptions regarding the dimensions of a
subyard of 120 m by 50 m and an angle of repose of 35° of the
stacked material, the capacity of a subpile is about 50 000 t. With
a pocket size of 300 t, the pile consists of 167 pockets or cells.
Depending on the number of layers m, the number of cells in a
layer is n. The assignment of CV to each cell reflects the incoming
material flow and the up and down movement of the stacker. All
odd-numbered layers are stacked from left to right, and all
even-numbered layers are stacked from right to left. The
reclaiming process occurs column by column. Because of the
constant cross section area of the pile, the necessary length to be
reclaimed for a unit of 1500 t is constant. Depending on the length
of a cell in the pile, multiple columns can contribute to reclaim
one unit of 1500 t, or only a part of a column contributes to
reclaim one unit. The described model is implemented using VBA
for Excel and is used to investigate the homogenisation effect of
the coal stock and blending yard.
Advances in Orebody Modelling and Strategic Mine Planning I
 INVESTIGATING THE VARIABILITY OF KEY COAL QUALITY PARAMETERS IN CONTINUOUS MINING OPERATIONS

Width of the subyard

FIG 9 - Mode of operation of a blending yard of type strata.

Reclaiming

CVRmS1 CVRmS2 CVRmS... CVRmSn-1 CVRmSn

Stacking

CVRm-1S1 CVRm-1S2 CVRm-1S... CVRm-1Sn-1 CVRm-1Sn

Stacking
CVR...S1 CVR...S2 CVR...S... CVR...Sn-1 CVR...Sn
CVR2S1 CVR2S2 CVR2S... CVR2Sn-1 CVR2Sn
CVR1S1 CVR1S2 CVR1S... CVR1Sn-1 CVR1Sn
FIG 10 - Model of the coal stock and blending yard.

Analysis of the effect of homogenisation in
dependence on the number of stacked layers
The following analysis is based on the results of the prediction of
CV, when operating two excavators (operating mode 3). The
application of the 25 realisations of the CV of the incoming
material flow to the previously described model of the blending
yard results in 25 realisations of the CV of the outgoing material
flow. Figure 11 shows the incoming material flow, while Figure 12
shows the outgoing material flow for the case of 15 layers in the
pile. The figures show the predicted CV per 1500 t unit for each
realisation. In addition, the average of the realisations and a 90 per
cent confidence level is shown.
There is a clear difference between incoming and outgoing
CV, which is the effect of the homogenisation of the blending
yard. Two effects can be detected:
1. the blending yard strongly transforms the varying CV of the
incoming material flow to a very homogeneous flow and
hardly disperses the CV of the outgoing material flow, and
2. a strong improvement of the certainty of predicting the CV.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 171
J BENNDORF

Predicted coal quality (CV) of the in-coming material flow similar homogenisation effect. This has implications on the
14,00 operation mode of the blending yard. The number of stacking
13,00
passes can be reduced from 61 to 35, which cause a reduction of
operational costs and increases the safety level in the operating
12,00
area. For the improvement of predicting CV, a similar observation
11,00 can be done. The standard deviation of units of the outgoing
Calorific value (CV) in MJ/kg

material flow is about 50 per cent of the incoming material flow.

10,00
This effect can also be achieved with 35 layers. The average
9,00 absolute value of the standard deviation of predicted CV of the
8,00
outgoing 1500 t units is ± 0.30 to ± 0.35 MJ/Kg. This represents
the standard deviation of predicting a block of 50 000 t in the
7,00
deposit that means a block of the size of a subpile in the blending
6,00 yard. In general, the standard deviation of predicted CV of the
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34
Extrated units of 1.500t
1500 t units in the outgoing material flow converges to the
standard deviation of the predicted CV of the whole subpile. This
FIG 11 - Predicted CV of the incoming 1500 t coal units.
observation is essential in designing subpiles and defining its size
as a function of the exploration level of the deposit and the
customer’s specifications.
Predicted coal quality (CV) of the in-coming material flow (after reclaiming)
14,00

13,00
SUMMARY AND CONCLUSIONS
12,00 The properties of the conditional simulation in geostatistics allow
11,00
for investigating the in situ variability of coal quality parameters
Calorific value (CV) in MJ/kg

and its behaviour along the extracting, conveying and blending

10,00
9,00
8,00
7,00
6,00
1 3
FIG 12 - Predicted CV of the outgoing 1500 t coal units.
The homogenisation effect of the blending yard is due to the
elimination of successive auto-correlated units. It is highly
probable that a unit of high CV is followed by another unit of a
high CV in case of the incoming material flow. This correlation
is disconnected in the outgoing material flow, which is caused by
the fan-shaped reclamation of the layered stacked coal.
To investigate the homogenisation effect, the above study is
performed on a varying number of layers in the pile. The number
takes values in an interval from one to 61, which defines the
technical limits of the blending yard under consideration. Results
in terms of the homogenisation effect and improvement of
certainty in predicting CV of shipped 1500 t units are presented
in Figure 13.
Effect of homogenisation and improvement in certainty in prediction
1,00
Improvement in prediction
process. The case study illustrated an application in a continuous
mining environment with different types of excavators and the
utilisation of a coal stock and blending yard. The results deliver
valuable insight into the effect of homogenisation and an
improvement in predicting CV values of shipped coal units as a
function of the operational mode. Different excavators cause a
5 7 9 11 13 15 17 19 21 23
Reclaimed units of 1500t P(train loads)
25 27 29 31 33 different behaviour of CV of the extracted coal. Bucketwheel
excavators cause substantial differences between extracted units,
which show a cyclic behaviour. Chain excavators tend to smooth
out CV and show a homogenisation effect. When operating both
excavators simultaneously, an improvement in predicting CV can
be achieved. Furthermore, the advantage of utilising a strata-type
stockpile in the conveying process was demonstrated by a
substantial improvement in predicting and homogenisation CV.
With a goal of meeting production goals with a high probability,
as well as delivering a homogeneous product, an operation of
excavators without utilising the stockpile should be avoided. The
results are a basis for optimising the operation and adjusting the
operational mode day-to-day, depending upon the customer’s
requirements and the current geological setting.
REFERENCES
Benndorf, J, 2009. Evaluation of lignite deposits using conditional
simulation in geostatistics, Schriftenreihe des Institutes für
Geotechnik und Markscheidewesen der TU Clausthal, Heft 18/2009,
221 p (in German).
20,0 Benndorf, J and Dimitrakopoulos, R, 2010. Stochastic long-term
18,0 production scheduling of iron ore deposits – Integrating joint
16,0 multielement geological uncertainty, in Advances in Orebody

Modelling and Strategic Mine Planning I (ed: R Dimitrakopoulos),

effect of homogenization

0,80
14,0
pp 151-158 (The Australasian Institute of Mining and Metallurgy:
12,0
0,60 Melbourne).
10,0
Costa, J F, Zingano, A C and Koppe, J C, 2000. Simulation – An
8,0
0,40 approach to risk analysis in coal mining, Exploration and Mining
6,0
Geology, 9(1)43-49.
4,0
0,20
Improvement in prediction Effect of homogenization
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation
2,0
(Oxford University Press: New York).
0,00 0,0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 35 41 45 51 61
Hohn, M E and McDowell, R R, 2001. Stochastic simulation of coal bed
Number of stacked layers thickness and economic decision-making, Geologic Modelling and
Simulation: Sedimentary Systems, (eds: D F Merriam and J C Davis),
pp 271-283 (Kluwer Academic/Plenum Publishers: New York).
FIG 13 - Homogenisation effect and improvement of certainty in
MIBRAG, 2009. MIBRAG web site [online]. Available from: <www.
predicting CV as a function of layers in the blending yard.
mibrag.de> [Accessed: September 2009].
Pavloudakis, F and Agioutantis, Z, 2001. Development of a software tool
In the case study, the dispersion between the delivered coal units for the prediction of the coal blending efficiency in longitudinal
can be reduced to a 17th of the incoming material flow by utilising stockpiles, in IMCET 2001, 17th International Mining Congress and
all 61 layers. However, 35 layers appear to be sufficient for a Exhibition of Turkey, Ankara, Turkey, 19 - 22 June.

172 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I

Cut-Off Grade Estimation – Old Principles Revisited – Application
to Optimisation of Net Present Value and Internal Rate of Return
J-M Rendu1
ABSTRACT
The modern principles of cut-off grade estimation were first published by
Ken Lane over 40 years ago. These principles form the basis of most
computerised systems routinely used in mine design and production
scheduling. In this paper, the author summarises these principles and
proposes simple equations to estimate cut-off grades in a variety of
situations. All consequences of choosing a cut-off grade are discussed
including financial, technical, economic, legal, environmental, social and
political ones. A utility function is proposed as a means to take into
account various potentially conflicting objectives. An example is given
which illustrates how cut-off grades can be calculated to maximise Net
Present Value (NPV) and Internal Rate of Return (IRR). The complex
relationship between naturally fixed geologic parameters and parameters
on which humans have some control is discussed.
INTRODUCTION
A cut-off grade is generally understood to be the value of a
geologically defined parameter used to separate material which
should be wasted from that to be sent to the processing plant. In
the simplest situations, the cut-off grade is a minimum metal
content below which material is sent to the waste dump and
above which it is sent to the mill. Cut-off grades are also used to
decide the routing of mined material when two or more processes
are available, such as heap leaching and milling. Cut-off grades
are used to decide whether material should be stockpiled for
future processing or processed immediately. Many publications
can be found in the technical literature concerning estimation
and optimisation of cut-off grades, the most comprehensive
being the reference book by K F Lane, The Economic Definition
of Ore: Cut-off Grades in Theory and Practice (Lane, 1988). An
introduction to the subject can be found in Rendu (2008).
This paper attempts to clarify the many consequences of
changing cut-off grades. Cut-off grades are calculated by
comparing costs and benefits. A cut-off grade strategy defines
the properties of the material being sent to the processing plant,
including tonnage, average grade and metallurgical properties.
Expected cash flows are function of this strategy. Most cut-off
grade optimisation algorithms are based on the assumption that
the owner’s objective is to maximise the NPV of future cash
flows, NPVi calculated using a specified discount rate i.
To reach this objective one must consider two classes of
variables:
• Class A – relating to the properties of the naturally occurring
mineralisation, and
• Class B – relating to human-driven economic constructs and
technology.
Class A variables are defined by nature and include the shape
of the deposit and its geologic and metallurgical properties. Class
B variables are at least partly the result of engineering, economic
and other factors that control mining sequence, costs, revenues
from sales and ultimately cash flows. The complex relationship
between Class A and Class B variables must be taken into
account to optimise cut-off grades. The complexity of cut-off
1. FAusIMM, Mining Consultant, 5451 South Geneva Street, Englewood
CO 80111, USA. Email: JMRendu@aol.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
grade optimisation is unavoidable. Decisions must be made using
complex objectives or combinations of potentially conflicting
objectives. Fundamental principles are first revisited, which give
a framework within which quantifiable and occasionally not
quantifiable variables can be taken into account. How these
principles can be used to make increasingly complex decisions is
illustrated with financial objectives which include optimisation
of cash flow, of Net Present Value and of Internal Rate of Return.
UTILITY FUNCTION
Choosing a cut-off grade is equivalent to choosing the value xc of
a geologically defined parameter (or set of parameters) x which
will be used to decide whether one tonne of material that becomes
available for processing in a particular period of time should be
sent to one process or another. In simple cases, a single parameter
may be sufficient to define the destination of material being
mined, such as a copper grade or a gold grade. In other cases, a set
of parameters may have to be considered such as copper and
molybdenum grades, sulfide content, clay content, mineralogy and
percentage of deleterious elements and the cut-off ‘grade’ must be
defined as a function of these parameters. The term ‘utility’ is used
here as it is in decision theory (Raiffa and Schlaifer, 1961) to
represent the satisfaction gained from following a given course of
action. This satisfaction is a function of preferences and values
specific to the decision maker. The utility of a given cut-off grade
strategy is a measure of the extent to which this strategy reaches
the mining company’s objectives. The utility of sending one tonne
of material with parameter value (grade) x to destination 1
(Process 1) is U1(x). The utility of sending the same material to
destination 2 (Process 2) is U2(x). The cut-off grade xc is the value
of x for which:
U1(xc) = U2(xc)
If U1(x) exceeds U2(x) for x greater than xc then all material
for which x is greater than xc should be sent to Process 1. The
choice of a cut-off grade is governed primarily by financial
objectives. However the consequences of choosing a given
cut-off grade are complex and not all of a financial nature. A
variety of possibly conflicting objectives must be taken into
account. To illustrate the role of these different objectives in the
decision making process, the utility function U(x) is written as
the sum of three parts:
U(x) = Udir(x) + Uopp(x) + Uoth(x)
Udir(x) represents the direct, undiscounted profit or loss which
will result from processing one tonne of material of grade x.
Uopp(x) represents the opportunity cost or benefit of changing the
processing schedule by adding one tonne of grade x to the
material flow. When there are capacity constraints, adding a new
tonne of material to a predetermined processing schedule cannot
be done without changing this schedule. Such a change may have
positive or negative consequences which are incorporated in the
opportunity cost. Opportunity costs occur only when there are
constraints, which limit how many tonnes can be mined or
processed at a given time. Other factors that must be taken into
account in the calculation of cut-off grades, but may not be
quantifiable, are represented by Uoth(x).
Spectrum Series Volume 17 173
J-M RENDU
DIRECT PROFIT AND LOSS
The marginal cut-off grade is calculated taking only direct
operating costs and corresponding revenues from sales into
account, if there are any. Capacity constraints are ignored. Cash
flows are not discounted. Opportunity costs are not taken into
consideration, and neither are any other consequences which
changing the cut-off grade may have on the economics of the
project. Only costs that can be directly attributed to the mining
and/or processing of one tonne of material are taken into
account. This may include some costs classified as overhead, to
the extent that such costs will change with the life or size of the
operation. The marginal cut-off grade is calculated by reducing
the utility function to its first term:
U(x) = Udir(x)
The utility of mining and processing one tonne of material of
grade x is:
Uore(x) = x × r × (V - R) – (Mo + Po + Oo)
where:
x represents the average grade
r is the proportion of valuable product recovered from the
mined material
V represents the value of one unit of valuable product
R stands for refining costs that are related to each unit of
valuable material produced
Mo represents the mining cost per one tonne of ore
Po the processing cost per one tonne of ore
Oo the overhead cost per one tonne of ore
The utility of mining and wasting the same material is:
Uwaste(x) = - (Mw + Pw + Ow)
where:
Mw represents the mining cost per one tonne of waste
Pw represents the processing cost per one tonne of waste that
may be needed to avoid potential water contamination and
acid generation
Ow stands for the overhead cost per one tonne of waste
To decide whether one tonne of material should be wasted or
sent to the processing plant, one must compare the utility of both
actions. The minimum cut-off grade is the value xc of x for
which:
Uore(xc) = Uwaste(xc)
xc = [(Mo - Mw) + (Po - Pw) + (Oo - Ow)]/[r × (V - R)]
In this formula, the numerator represents the difference
between direct costs incurred when treating the material as ore
and those incurred when treating it as waste. The denominator
represents the revenues resulting from the sale of one unit of
contained valuable product. If the costs of mining and shipping
material to the waste dump or to the processing plant are the
same (Mo = Mw) and there are no significant additional costs in
processing waste (Pw = 0 and Ow = 0), this cut-off grade is only a
function of mill costs and recoveries and is independent of
mining costs. It is known as the mill cut-off grades:
xc = [Po + Oo]/[r × (V - R)]
If the material considered need not be mined because it is at
the bottom of the pit or along the walls of a stope, the utility of
treating it as waste is Uwaste(x) = 0. The material will be mined
and processed only if Uore(x) is positive. This cut-off grade is a
174 Spectrum Series Volume 17
function of both mining and milling costs and is known as the
mine cut-off grades:
xc = [Mo + Po + Oo]/[r × (V - R)]
OPPORTUNITY COSTS AND BENEFITS
Opportunity costs Uopp(x) may result from mining and processing
one tonne of material that was not previously scheduled to be
mined or processed. No opportunity cost is incurred if mine, mill
or refining facilities are not capacity constrained – only direct
costs and benefits need be taken into account. If there are
capacity constraints, adding a new tonne to the production
schedule will result in displacing material already scheduled to
be processed and postponing treatment of this material. The cost
of displacing this material is the opportunity cost. The
opportunity cost must be added to the direct cost of the process
that is capacity-constrained. If the mine is capacity-constrained,
the opportunity cost applies to both ore and waste and has no
impact on cut-off grade, unless waste material can be left in the
ground, as may be the case at the bottom of an open pit or along
the walls of an underground stope. If the mill is capacity-
constrained the opportunity cost is added to all tonnes processed.
Consider a project whose production schedule has been
designed to fill up the processing plant to capacity. The Net
Present Value of future cash flows is NPVi. This Net Present
Value was calculated on the basis of currently planned
production using a discount rate i. If we add one new tonne of
material to the capacity-constrained processing plant we are
postponing treatment of the originally scheduled material by the
time t needed to process the additional tonne. This time is equal
to the inverse of the plant capacity. Adding one new tonne today
will decrease the Net Present Value of previously scheduled
future cash flows by t × i × NPVi. The opportunity cost of adding
one tonne of material to a capacity-constrained operation is:
Uopp(x) = - t × i × NPVI
Because the Net Present Value of future cash flows decreases
over time, the opportunity cost also decreases and the optimal
cut-off grade increases. At the end of the mine life the
opportunity cost is zero and the applicable cut-off grade is the
minimum cut-off grade calculated using only direct costs. The
above formula is useful to verify that cut-off grades and Net
Present Value have been optimised. However cut-off grades
calculated from cash flows which have not been optimised are
also not optimal and an iterative approach must be used to
estimate the cut-off grade. For example one can first calculate a
cash flow using a fixed cut-off grade such as that calculated
without opportunity cost. The Net Present Value of this cash flow
can be calculated and the theoretically optimal cut-off grade can
be re-estimated using the corresponding opportunity cost. But
new cut-off grades imply new mine plans, new cash flows, and
therefore new opportunity costs. These updated opportunity costs
must be used to re-estimate the cut-off grades. If the theoretically
optimal cut-off grades are significantly different from those
which were used in the Net Present Value calculation, significant
successive modifications to the mine plan may be needed before
an optimum solution is found.
Algorithms can be found in the technical literature and
computer programs have been developed to facilitate cut-off grade
optimisation. Because of the complex relationship between the
geological properties of the deposit and human-driven parameters
which define production schedule and cash flow, no simple
solution exists while improved algorithms and easier to use
computer programs are continuously developed. Examples of
proposed solutions can be found in publications by Carter, Lee,
and Baarsma (2004), Chanda (2004), Dagdelen and Kawahata
(2008), Dimitrakopoulos, Martinez and Ramazan (2004), Hall
Advances in Orebody Modelling and Strategic Mine Planning I

(2003), Hoerger, Bachmann, Criss, and Shortridge (1999),
Hoerger, Hoffman, and Seymour (1999), King (2001), Lerchs and
Grossmann (1965), Lane (1988), Whittle (1999) and others.
OTHER COSTS AND BENEFITS
Cut-off grades play a critical role in defining all major
characteristics of a mine, including the size of operation, capital
and operating costs, life of the mine and cash flow. In addition to
the economically quantifiable financial impact that cut-off grade
changes may have, there are other costs and benefits which are
often not easily quantifiable but must be taken into account.
Consideration must be given not only to changes in cash flow and
Net Present Value as measured by Udir(x) and Uopp(x), but also to
all other impacts Uoth(x), including those of an environmental,
socio-economic, ethical and political nature. All stakeholders must
be taken into account to evaluate different cut-off grade strategies.
These stakeholders include:
• shareholders;
• banks and other financial institutions;
• regulators;
• employees and their families;
• customers who buy the product sold by the company;
• suppliers of goods and services;
• local communities;
• local, regional, federal and country governments;
• non-governmental agencies (NGOs); and
• future generations which will live with the long-term
environmental impact of the mining operation.
Management decides how to balance financial goals with other
stakeholder requirements and expectations. The success of a
project and the company’s long-term public image and licence to
operate depend on skilful management of potentially conflicting
objectives. Practical guidelines must be given to those in charge
of mine planning to ensure that the projects are designed to reach
the company’s objectives.
Higher cut-off grades may increase short term profitability,
enhance return to shareholders, shorten the payback period and
reduce political risk. However, for potentially long-life projects,
significantly shortened mine life can reduce time-dependent
opportunities such as those offered by price cycles. Lower cut-off
grades may increase project life, with potentially more
sustainable economic impact. Lower cut-off grades imply fuller
consumption of mineral resources, which may be politically
advantageous or legally required. However, lower cut-off grades
may also result in higher sensitivity to price and cost variations.
Stakeholders may have to choose between higher returns over
short time periods or low returns over longer time periods. In
some circumstances, using high but decreasing cut-off grades
early in the mine life and stockpiling low-grade material for later
processing can help balance economic uncertainty, financial
returns, resource consumption and mine life.
A method that can be used to optimise cut-off grades while
taking into account costs and benefits which cannot be
quantified, consists of evaluating the project under a variety of
constraints imposed on mine capacity, mill capacity, volume of
sales, capital or operating costs, life of mine, etc. A different
cut-off grade profile is developed under each set of constraints.
These cut-off grades are chosen to optimise Net Present Value or
Internal Rate of Return. The relationship between discount rate
and Net Present Value, the project Internal Rate of Return, its
sensitivity to cost and price changes and risk analysis, are some
of the measures used to compare projects. The final development
decision is made by comparing the quantified advantages and
Advances in Orebody Modelling and Strategic Mine Planning I
CUT-OFF GRADE ESTIMATION – OLD PRINCIPLES REVISITED
disadvantages of each project with the subjective cost/benefit
to all stakeholders, according to company-specific values and
priorities.
DIFFERENCE BETWEEN MAXIMISING NPV
AND MAXIMISING IRR
The choice between business objectives is a rational decision
based on the environment within which the company operates,
including the availability of mutually exclusive and/or independent
investment opportunities and the extent to which capital resources
are constrained. Optimisation of Net Present Value is the objective
most commonly used to calculate cut-off grades. In some
situations management’s primary objective might be maximisation
of the Internal Rate of Return (IRR).
Consider the case where a feasibility study has been completed
on a new project. The capital cost of bringing the project to
production is I. Production schedules have been developed and
corresponding cash flows have been estimated. Using a discount
rate i, the Net Present Value of expected future cash flows is
NPVi. This Net Present Value is a function of the point in time
when it is calculated. By definition the Internal Rate of Return
IRR is the value ir of i for which the Net Present Value of the
project, calculated before the investment is made and therefore
including the capital investment I, is zero:
NPVir = 0
If there was no capacity constraint, maximum rate of return and
maximum Net Present Value would both be obtained by
maximising production over the shortest possible time period. In
practice, except for occasional periods during the life of a mine, or
when mine and mill have been over-designed, all mining
operations are capacity constrained. Such constraints are imposed
by the need to balance capital cost, operating cost and production
schedule. Without capacity constraint, the optimum cut-off grade
xc1 between material wasted and that sent to the mill is:
xc1 = [(Mo - Mw) + (Po - Pw) + (Oo - Ow)]/[r × (V - R)]
If the mill is capacity-constrained, adding one tonne to the
production schedule results in an opportunity cost Uopp(x):
Uopp(x) = - t × i × NPVi
A higher cut-off grade must be used, xc2, calculated as follows:
xc2 = [(Mo - Mw) + (Po - Pw) + (Oo - Ow) + t × i × NPVi]/
[r × (V - R)]
This is the optimum cut-off grade when the primary objective is
to maximise the Net Present Value using the discount rate i. In this
formula, the Net Present Value NPVi must be calculated from
expected future cash flows, beginning with the period for which
the cut-off grade is calculated. The cut-off grade that applies to
year one must be calculated using the Net Present Value of cash
flows beginning with year one. This Net Present Value is a
function of the production schedule but is independent of any past
investment. Past investments enter into the calculation indirectly
and they enter only because of capacity constraints that were
imposed by investment decisions made at an earlier date. Because
the Net Present Value of future cash flows usually decreases when
the remaining mine life decreases, the cut-off grade also decreases
over time.
If maximisation of Internal Rate of Return is the primary
objective, the desired discount rate is ir as defined above. The
formula to be used to calculate the optimal cut-off grade can be
re-written as follows:
xc3 = [(Mo - Mw) + (Po - Pw) + (Oo- Ow) + t × ir × NPVir]/
[r × (V - R)]
Spectrum Series Volume 17 175
J-M RENDU
In this formula, ir is the calculated Internal Rate of Return and
is a function of the capital investment I. At the beginning of the
first year of production the Net Present Value NPVir is equal to
the total initial investment I, updated to year one, taking into
account the discount rate i in the calculation of the present value
of past investments. The same formula applies to the calculation
of cut-off grade, whether the objective is to maximise NPV or
IRR, except that in the first case the discount rate is fixed while
in the second case it changes with the size of the initial
investment.
When considering a new project, different development options
must be considered. Different mine and mill capacities are
considered, resulting in different production schedules, cash flows,
and capital requirements. Each one of these options corresponds to
a different Internal Rate of Return and optimal cut-off grade
profile. For each option, the Internal Rate of Return is a function
of planned production, which in turn is determined by the cut-off
grade. Given an estimated cash flow, one can calculate a
theoretically optimal cut-off grade profile. If this theoretical
optimum is different from the cut-off grade actually used to
develop the production schedule, changes to the cut-off grade
should be considered.
Example
The following example illustrates how cut-off grades can be
estimated depending on whether the objective is maximisation of
Net Present Value or of Internal Rate of Return. A feasibility
study has been completed on a proposed new copper mining
operation. The capital investment needed to bring the mine to
production has been estimated at $1.50 billion. The planned
mining capacity is 72 million tonnes per year (including both ore
and waste), and the mill capacity is 36 million ore tonnes per
year. The following operating parameters have been estimated:
r = 86% (including mill and smelter recoveries)
V = $1.20 per pound of copper sold
R = $0.30 per pound of copper for freight, smelting and refining
Mo = $1.45 per tonne of ore mined
Po = $4.00 per tonne of ore processed
Oo = $0.50 per tonne of ore processed
Mw = $1.45 per tonne of waste
Pw = $0.05 per tonne of waste
Ow = $0.05 per tonne of waste
The feasibility study was based on a 14-year mine life and a
mill head grade of 0.91 per cent Cu. Using a ten per cent
discount rate, the project Net Present Value is $630 000 000 and
the Internal Rate of Return is 17.2 per cent. Figure 1 shows the
CashFlow NPV10% NPV17.2%
2.5E+09
NPVofFutureCashFlow
2.0E+09
1.5E+09
1.0E+09
5.0E+08
0.0E+00
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Year
FIG 1 - Yearly cash flow and Net Present Value of future cash flows.
176 Spectrum Series Volume 17
expected yearly cash flow and Net Present Value of future cash
flows discounted at ten per cent and 17.2 per cent respectively.
Two modifications to the mine plan are being considered – the
objectives being to optimise the Net Present Value at a ten per
cent discount rate and to maximise the Internal Rate of Return,
respectively. The only constraint imposed on these modifications
is that the mill capacity must remain at 36 million tonnes. A first
step consists in calculating the theoretically optimal cut-off grade
for Year 1. To do this, one calculates the Net Present Value of
future cash flows as of the beginning of Year 1:
• NPV10% = $2 130 000 000
• NPV17.2% = $1 500 000 000
If there was no capacity constraint the minimum cut-off grade
would be:
xc1 = [(Mo - Mw) + (Po - Pw) + (Oo - Ow)]/[r × (V - R)]
= [(1.45 - 1.45) + (4.00 - 0.05) + (0.50 - 0.05)]/[0.86 ×
(1.20 - 0.30) × 2205]
= 4.400/1707 = 0.26% Cu
In this formula, 2205 is the number of pounds per tonne. If the
objective is to maximise the Net Present Value NPV10% the Year 1
opportunity cost to be added to the processing cost is calculated
as follows:
Uopp(x) = - t × i × NPV10%
= - (1/36 000 000) × 10% × 2 130 000 000
= -$5.917 per tonne mined
The theoretically optimal Year 1 cut-off grade is xc2:
xc2 = (4.400 + 5.917)/1707 = 0.60% Cu
If the objective is to maximise the Internal Rate of Return the
Year 1 opportunity cost to be added to the processing cost is
calculated as follows:
Uopp(x) = - t × ir × I
= - (1/36 000 000) × 17.2% × 1 500 000 000
= -$7.167 per tonne mined
The theoretically optimal Year 1 cut-off grade is xc3:
xc3 = (4.400 + 7.167)/1707 = 0.68% Cu.
The cut-off grades xc2 and xc3 are optimal only during the first
year of production, provided that Net Present Value and Internal
Rate of Return were calculated from mine plans which used
these cut off grades. Figure 1 shows how the Net Present Value
of future cash flows decreases over time. Figure 2 shows how the
cut-off grade, calculated using these Net Present Values,
decreases over time.
This example illustrates why optimisation of cut-off grades is a
difficult, lengthy, iterative process. Cut-off grades calculated using
the above formulae are optimal only if the cash flows used to
calculate these cut-off grades were developed using optimal
cut-off grades. In a typical situation initial assumptions will be
made concerning the mine and plant capacities which are
technically feasible, consistent with geologic constraints and
expected to give a reasonable rate of return. Production schedules
will be developed using cut-off grades not lower than the marginal
cut-off grade. Capital costs and operating costs will be estimated
accordingly, from which cash flows will be developed. It is not
likely that the cut-off grades used in this initial production
schedule and cash flow analysis will be optimal. To reach
the optimum, one must consider changing the production
schedule, taking into account the calculated theoretically optimum
cut-off grades. These changes are subject to geologic, technical
Advances in Orebody Modelling and Strategic Mine Planning I

Nodiscount 10.0%discount 17.2%discount
0.8%
0.7%
CutoffGrade(Cu%)
0.6%
0.5%
0.4%
0.3%
0.2%
0.1%
0.0%
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Year
FIG 2 - Yearly cut-off grades at various discount rates.
and financial constraints. In the example above, supplying
36 000 000 t to the processing plant may not be achievable if the
cut-off grade is raised to 0.68 per cent Cu. Significant changes in
cut-off grade may require commensurate changes in mine
production schedules. Mining constraints may be encountered
before the necessary tonnage of mill feed can be mined. These
constraints must be taken into account in addition to the
processing constraints. When both mine and mill are capacity
constrained, cut-off grades are determined by the geologic
properties of the deposit, including tonnage and grade of material
which can be mined. Financial considerations become secondary.
CONCLUSIONS
The formulae presented in this paper are simple and easy to use.
They are particularly useful to check whether the cut-off grades
planned for development of a new mine, or actually used in an
operating mine are optimal or nearly so. They are also useful to
check whether production schedules developed using available
mine planning and scheduling software are reasonable. However,
geologic environment and changes in operating and economic
conditions can impose constraints which require a cut-off grade
significantly different from the theoretical optimum at any time.
This difference should be analysed to make sure that it can be
explained and justified. A theoretical optimum is only an
optimum if all constraints are taken into account. It is because of
the complex relationship between fixed parameters, determined
by geology and parameters on which mine operators have some
control, that cut-off grade optimisation remains an elusive
problem which will keep many practitioners and researchers
challenged for years to come.
Advances in Orebody Modelling and Strategic Mine Planning I
CUT-OFF GRADE ESTIMATION – OLD PRINCIPLES REVISITED
REFERENCES
Carter, P G, Lee, D H and Baarsma, H, 2004. Optimisation methods for
the selection of an underground mining method, in Proceedings
Orebody Modelling and Strategic Mine Planning (ed: R
Dimitrakopoulos), pp 7-12 (The Australasian Institute of Mining and
Metallurgy: Melbourne).
Chanda, E K, 2007. Network linear programming optimisation of an
integrated mining and metallurgical complex, in Orebody
Modelling and Strategic Mine Planning, second edition (ed: R
Dimitrakopoulos), pp 149-155 (The Australasian Institute of Mining
and Metallurgy: Melbourne).
Dagdelen, K and Kawahata, K, 2008. Value creation through strategic
mine planning and cutoff-grade optimisation, Mining Engineering,
60:39-45.
Dimitrakopoulos, R, Martinez, L and Ramazan, S, 2007. Optimising open
pit design with simulated orebodies and Whittle Four-X: A
maximum upside/minimum downside approach, in Orebody
Modelling and Strategic Mine Planning, second edition (ed: R
Dimitrakopoulos), pp 201-206 (The Australasian Institute of Mining
and Metallurgy: Melbourne).
Hall, B E, 2003. How mining companies improve share price by
destroying shareholder value, in Proceedings CIM Mining
Conference and Exhibition, paper 1194, (The Canadian Institute of
Mining and Metallurgy: Montreal).
Hoerger, S, Bachmann, J, Criss, K and Shortridge, E, 1999. Long-term
mine and process scheduling at Newmont’s Nevada operations, in
Proceedings 28th APCOM Symposium (ed: K Dagdelen), pp 739-747
(The Society for Mining, Metallurgy and Exploration Inc: Littleton).
Hoerger, S, Hoffman, L and Seymour, F, 1999. Mine planning at
Newmont’s Nevada operations, Mining Engineering, 51:26-30.
King, B, 2001. Optimal mine scheduling, mineral resource and ore
reserve estimation, in Mineral Resource and Ore Reserve Estimation
– The AusIMM Guide to Good Practice (ed: A C Edwards),
pp 451-458 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Lane, K.F, 1988. The Economic Definition of Ore: Cut Off Grades in
Theory and Practice, 149 p (Mining Journal Books Ltd: London).
Lerchs, H and Grossmann, I F, 1965. Optimum design of open-pit mines,
Transactions, LXVIII:17-24 (The Canadian Institute of Mining and
Metallurgy: Montreal).
Raiffa, H and Schlaifer, R, 1961. Applied Statistical Decision Theory,
356 p (Harvard University: Boston).
Rendu, J-M, 2008. An Introduction to Cutoff Grade Estimation, 112 p
(The Society for Mining, Metallurgy and Exploration Inc: Littleton)
Whittle, J, 1999. A decade of open pit mine planning and optimisation –
The craft of turning algorithms into packages, in Proceedings 28th
APCOM Symposium (ed: K Dagdelen), pp 15-23 (The Society for
Mining, Metallurgy and Exploration Inc: Littleton).
Spectrum Series Volume 17 177

Using Grade Uncertainty to Quantify Risk in the Ultimate Pit
Design for the Sadiola Deep Sulfide Prefeasibility Project, Mali,
West Africa
S Robins1
ABSTRACT
In order to quantify the uncertainty in the grade estimate for the Sadiola
deep sulfide prefeasibility project, a conditional simulation model was
generated using the direct block simulation methodology. Compared to
conventional sequential Gaussian simulation, the direct block simulation
algorithm produced a reliable model in significantly less time, lending its
application to a production environment. Through application of a mining
transfer function, risk pits were generated for comparison with the deep
sulfide prefeasibility pit. The results of this study revealed that the
prefeasibility pit is optimal at the applied gold price and cost parameters,
and that the risk of not achieving the project grade profile is low. Should
the gold price increase, or the operating costs of the project decrease
significantly, the deep sulfide reserve tonnage would realise significant
upside potential. Probability and uncertainty analysis revealed that the
greatest risk to the project is the confidence in the footwall grade
estimate. At a drill spacing of 50 m × 50 m and a sample interval of 1 m,
the probability of the footwall grade exceeding the economic cut-off of
2.0 g/t is low, while the uncertainty in the grade estimate is high.
Although significantly lower in grade than the main zone, which is the
primary economic driver of the project, the footwall mineralisation is
important in terms of reducing stripping ratio and delivering ore tonnes to
optimise the treatment schedule. This zone is therefore a focus area for
further drilling.
INTRODUCTION
Sadiola Hill Gold Mine is located at latitude 13°56’N, longitude
11°40’W, and altitude 125 m above mean sea level, which places
it approximately 500 km north west of Bamako, the capital of
Mali in West Africa (Figure 1). The mine is operated by the
Societe d’Exploration des Mines d’Or de Sadiola SA (SEMOS
SA), which comprises a joint venture partnership between
AngloGold Ashanti (38 per cent), IAMGOLD (38 per cent), the
Malian Government (18 per cent) and the IFC (six per cent).
Construction of the Sadiola plant began in March 1995 and was
completed in June 1997. Gold production began in mid 1997
from the saprolite oxide orebody and more recently from
saprolite sulfide ore as mining has progressed through the
weathered profile towards the hard/soft boundary. The Sadiola
orebody extends substantially below the weathered horizons into
the underlying hard sulfide lithologies. However, the current Life
of Mine plan (LOM) does not exploit this material because the
plant is unable to treat more than ten per cent hard material. In
2005, Sadiola treated approximately 5.3 Mt of ore at 2.83 g/t,
accounting for a gold production of 483 koz (Van der
Westhuizen, 2005). The Sadiola plant was designed to treat this
material, comprised almost entirely of saprolite and laterite. At a
cut-off grade of 1.42 g/t, the remaining Sadiola Main Pit Reserve
(at the end of 2005) was 21 Mt at 3.35 g/t, accounting for
2.3 Moz. This reserve would accommodate mining to the end of
2009.
The hard sulfide portion of the Sadiola main pit mineral
resource at the end of 2005 was 48.5 Mt at 2.25 g/t, accounting
for 3.5 Moz (at 0.70 g/t cut-off grade), providing large upside
potential for the mine. Should Sadiola be able to treat hard
sulfide material, mining could continue to the end of 2015 and
1. Geology Manager, Geita Gold Mine, PO Box 532, Geita, Tanzania.
Email: srobins@geitagold.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
the LOM extended to 2024. SEMOS SA therefore initiated the
deep sulfide prefeasibility project to optimise the exploitation of
the hard sulfide component of its mineral resource. The
prefeasibility project focused on the pit expansion potential of
the deposit, as well as the options available to upgrade the
current gold treatment plant to process the hard sulfide material.
As of June 2006, cost estimates indicated that a capital outlay of
approximately USD $145 million would be required to upgrade
the treatment plant to efficiently process hard sulfide material.
Given the large capital outlay required to commence with the
deep sulfide Project, it was necessary to quantify the technical
risk of exploiting the Sadiola hard sulfide resource.
Quantification of the grade uncertainty was vital to providing
focus areas that required additional drilling before a reliable
feasibility study grade model could be obtained. The Sadiola
mineral resource model used for the prefeasibility study was
comprised of a grade estimate, into 30 m × 30 m × 10 m blocks,
using ordinary kriging. This was followed by a change of support
calculation, using uniform conditioning, to an SMU support size
of 10 m × 10 m × 5 m. The sample information on which this
model was built was comprised predominantly of NQ (53 mm
diameter) half diamond core, with 4.5 inch RC precollars
through the soft material. The final model comprised
13 independent estimation domains to honour the stationarity
requirements for kriging.
This model effectively generated a single grade outcome for
planning purposes. Questions remained as to how accurate this
grade estimate was, and if there was a practical method for
determining the uncertainty in the grade estimate. Conditional
simulation was utilised to answer these questions. Conditional
simulation, a Monte Carlo-type simulation approach, generates
multiple and equally probable realisations that provide a model
of the spatial uncertainty of the grade in the in situ orebody
(Dimitrakopoulos, Farelly, and Godoy, 2002). The resultant
realisations are not only conditioned to the available sample data,
and therefore all reasonably match the same sample statistics, but
also reasonably duplicate the histogram and semi-variogram
model of the sample data (Goovaerts, 1997). Spatial features are
deemed ‘certain’ if they are present in most of the simulated
maps and ‘uncertain’ if seen on a few simulated maps.
According to Goovaerts (1997), generating alternative
realisations of the spatial distribution of an attribute is rarely the
goal. Rather these alternative realisations serve as the input to
other transfer functions, which in the open pit mining
environment would comprise a mining process such as a pit
optimisation algorithm. In assessing the uncertainty in the grade
estimate for the Sadiola mineral resource model, the goal was to
generate a number of different grade realisations and run each of
these realisations through a pit optimisation algorithm in order to
determine where the prefeasibility design pit ranked. The
challenge lay in selecting a simulation method that was practical
to use in an operating mine environment. Sequential Gaussian
simulation (or SGS) (Goovaerts, 1997), the industry norm at the
time, proved impractical because of the computing time involved
in generating the 50 realisations required per estimation zone.
The relatively new method of direct block simulation was
therefore selected.
Spectrum Series Volume 17 179
S ROBINS

MAURITANI

Segala MALI

SENEGAL

Sabodala Loulou
Yalea
Tabakoto
Bamako

GUINEA

FIG 1 - Locality map for Sadiola mine, Mali, West Africa.

Dimitrakopoulos, Farelly and Godoy (2002), proposed the Senegalo-Malian Shear

direct block simulation (DBSIM) method as a viable alternative
to conventional simulation methods that considerably reduces
computational time. DBSIM detailed in Godoy (2003) and also
outlined in Boucher and Dimitrakopoulos (2009) and Peattie YATELA
(2005) simulates the internal points of each block, and when the SADIOLA
simulated block is calculated, the point values are discarded. The
simulated block value is then added to the conditioning dataset.
To integrate the block support conditioning data, the algorithm
has been developed in terms of a joint simulation, where the
second variable relates to the block values sequentially derived
through the simulation process. The algorithm simulates several SABODALA
LOULOU
hundreds of blocks per second and is considerably faster than YALEA
any point conditional simulation combined with reblocking. SEGALA
Furthermore, Godoy (2003), Benndorf and Dimitrakopoulos TABAKOTO
(2007) and others show that in addition to being substantially KENIEBA
faster and more efficient in terms of computing requirements, the
DBSIM method is reliable in terms of reproduction of the sample
statistics.

GENERAL GEOLOGY MEDINANDI

The Sadiola deposit is located within the Malian portion of the
Kenieba-Kedougou window, a major Early Proterozoic –
Birimian outlier along the NE margin of the Kenema – Man
shield (Figure 2). The Birimian of the window can be interpreted
as a collage of at least two NS trending terrains of different
nature. To the west, an older (±2.2 Ga) tholeitic mafic volcanic
unit with island arc type volcanics, intruded by a major
Keniebandi Saboussire Granitoids
calc-alkaline batholith, belongs to the Saboussire Formation. It is
separated from the dominantly sedimentary sequence of the Kofi Diale Dalema- Faults and
Formation by a major north-to-northeast trending shear zone. Kofi Shears
This sedimentary domain is significantly younger and is intruded
by calc-alkaline batholiths dated at 2.0 - 2.05 Ga. Within this FIG 2 - Geology of the Kenieba-Kedougou window.

180 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 USING GRADE UNCERTAINTY TO QUANTIFY RISK IN THE ULTIMATE PIT DESIGN
domain metamorphic grade is greenschist facies, with formation
of metamorphic biotite and locally amphibolite grade near major
intrusions. The Kofi Formation is obliquely cut by the
approximately N-S to N10° trending Senegalo-Malian shear zone
(SM), which is punctuated by several gold deposits along its
splays (Loulo, Yalea, Sadiola, Yatela), (Robins et al, 2005).
The Sadiola deposit is located in the north of the window and
is hosted by sediments of the Kofi Formation, which have been
intruded by numerous felsic intrusives. The dominant sediments
consist of fine-grained greywacke and impure carbonates with
minor tuffs and acid volcanics. At Sadiola, the intensely folded
impure carbonate packages comprise an alternation of limestone
beds, a few millimetres to several decimetres thick, with thinner,
more detritic beds. The Sadiola deposit occurs along the N10°
striking Sadiola Fracture Zone (SFZ), which is thought to be a
brittle-ductile splay off the SM Shear at a sinistral releasing
bend. The SFZ follows the steeply westerly dipping contact
between greywacke to the west and impure carbonate to the east.
Along the SFZ, both the greywacke and impure carbonate are
transposed. The SFZ and its wall-rock are injected by
discontinuous diorite dikes. Silicified quartz-feldspar-porphyry
(QFP) dikes often intrude along later steep west dipping, N20°
striking structures. The QFPs cross-cut the diorites and may
show brittle fracturing. Post mineralisation deformation has
complicated structural relationships in the deposit.
Gold mineralisation at Sadiola occurs along the SFZ over a
drilled strike length of approximately 2500 m, and remains open
to the north and south. N20° trending fault splays off the SFZ are
also well mineralised. The mine geological and grade block
model indicates the presence of 20° to 25° south dipping ore
shoots within the plane of the SFZ. Mineralisation occurs in all
of the four major rock types (marble, greywacke, diorite, and
FIG 3 - Section 5600N through Sadiola orebody. Mineralisation sits along the greywacke-marble contact (Sadiola Fracture Zone).
N20º faults are indicated in black. Section compiled by Dr A Smeesters, 2006.
Advances in Orebody Modelling and Strategic Mine Planning I
quartz-feldspar porphyry) and is spatially associated with a
complex alteration pattern. Drilling of the (unweathered) primary
mineralisation has allowed detailed investigation of major and
minor hydrothermal alteration processes that were active during
the formation of the deposit. Alteration assemblages identified to
date include calc-silicate, potassic, chlorite-calcite, carbonate
and silicification, and have allowed the deposit to be classified as
a mesothermal-shear-zone hosted deposit. Gold is associated
with both arsenic and antimony dominated sulfide assemblages
including arsenopyrite, pyrrhotite, pyrite, stibnite and gudmuntite
(Robins et al, 2005). Deposits of this type world-wide exhibit
good continuity of mineralisation along the strike and extending
to great depth. Structurally controlled, high grade pay shoots’
typically occur within a more pervasive lower grade halo. At
Sadiola, the location and geometry of high grade mineralisation
appears to be controlled by the confluence of the SFZ with the
N20° splays, resulting in steeply to vertically plunging zones
within the plan of the SFZ. The geometry of the extensive soft
‘saprolitic’ oxide deposit at Sadiola relates almost exclusively to
the supergene weathering history of the primary mineralisation.
The permeability of the rock formation, controlled mainly by
faulting, shearing and porosity, allows deep penetration of
ground water, causing oxidation of primary sulfides. Oxidation
of pyrite (and other sulfide species) results in the formation of
sulfuric acid, further promoting the downward argillisation of the
bedrock to form the clay rich assemblages present in the
saprolite. The irregular, ‘karst like’ soft rock/hard rock contacts
can be related to the extent of faulting and the original sulfide
content of the overlying profile. The intense weathering has
resulted in a tropical climate forming a series of decarbonated-
argillised troughs of variable depth (up to 180 m depth) along the
SFZ. This is the rich oxide orebody currently being mined
(Figure 3).
Spectrum Series Volume 17 181
S ROBINS

CREATION OF THE SIMULATION MODEL with the conditioning data zoned according to the estimation
domains (ZONECODE), outlined in Table 2. The input data
comprised raw grade information that was converted to normal
Simulation methodology space by the DBSIM program during the simulation runs. During
Although the deep sulfide project has focused specifically on the this process, DBSIM declustered the data, and output ‘equal
hard sulfide component of the Sadiola resource, a significant weighted’ statistics for the conditioning data, simulated nodes
amount of mineralised saprolite material occurs between the and blocks.
hard/soft boundary and the current LOM pit design. Since this
material would contribute towards the economics of the project, TABLE 2
it was necessary to account for it in the simulation model. The Table of estimation domains and their description.
laterite material only occurs close to the original ground surface,
and would have limited influence on the project economics since ZONECODE Description Project
most of the ore within this zone has already been mined. To 1000 Far north saprolite ore Main pit – oxide
speed up the simulation process, the laterite estimation domains
were excluded from the study. All other estimation domains used 2000 Saprolite waste Main pit – oxide
in the Sadiola recoverable resource model were honoured. 3000 North saprolite ore Main pit – oxide
Furthermore, the same densities as per the recoverable resource 4000 South saprolite ore Main pit – oxide
block model were retained in the simulation model (Table 1).
The determination of the various estimation domains in the 5000 Main ore – high grade Deep (hard) sulfide
recoverable resource model was based on trend analyses and 6000 Main Oore – low grade Deep (hard) sulfide
sample statistics. This process is outlined in Robins et al (2005). 7000 Footwall/hangingwall ore Deep (hard) sulfide
8000 North hard ore Deep (hard) sulfide
TABLE 1 9000 Waste Deep (hard) sulfide
Table of rock types and their associated densities.

ROCKTYPE Description Density Normal score semi-variogram models

code (g/cm3)
1 Laterite and clay 1.97 (average) Double structured, normal score spherical semi-variogram
models were calculated per ZONECODE (Table 3). The search
2 Saprolite 1.80 (average)
distances are those used for simulation in the x, y, and z
3 Silicified oxide 2.57 (average) orientations. The general characteristics of the normal score
4 Saprolite sulfide 2.00 (average) variograms can be summarised as follows:
5 Hard sulfide 1.55 to 3.03 (kriged) • The northern and southern mineralisation is separated by a
6 Blast oxide 2.10 (average) WNW trending fault that appears to have rotated the strike of
the mineralisation so that to the north, the mineralisation
7 Blast sulfide 2.10 (average) trends 030, and to the south, it trends 000.
• Most of the hard sulfide mineralisation (zones 5000, 6000
Sample data and 7000) is situated in the south of the orebody and
therefore their variogram structures resemble that of the
Almost all the deep sulfide sample data comprised NQ diamond South Saprolite ore (zone 4000), for which the relatively low
core at 1 m sample intervals. However, since a significant nugget is attributable to supergene enrichment of the grade.
amount of saprolite material – which was drilled with 125 mm The main ore high- and low-grade variograms (zones 5000
RC method and at 2 m sample intervals – would be included in and 6000 respectively) are similar in all respects.
the Deep Sulfide Project, all conditioning (sample) data was
composited to 2 m sample intervals. To test for clustering, data • The Far North (zone 1000) and North Saprolite (zone 3000)
was declustered using a moving window method. It was found ore zones have similar structure and orientation trending 030,
that there was no significant spatial clustering present in the data in line with the change in strike of the mineralisation. The Far
and therefore declustering was unnecessary. This was likely a North saprolite has significantly more continuity in the y
result of domaining the data and therefore achieving a relatively direction, though it was necessary to reduce the search
regular sample grid per domain. A sample input file was supplied distance in this direction to minimise negative kriging weights.

TABLE 3
Normal score semi-variogram models by ZONECODE.

ZONECODE C0 C1 C2 Range 1 Range 2 Search distance Azimuth

/dip
X Y Z X Y Z X Y Z
A2 A1 A3 A5 A4 A6
1000 0.190 0.510 0.300 14 35 15 50 130 30 30 30 20 030/00
2000 0.258 0.450 0.284 42 30 28 170 270 55 30 30 10 170/00
3000 0.200 0.400 0.398 10 10 12 40 50 30 25 40 15 030/00
4000 0.150 0.370 0.478 9 9 9 65 65 45 33 45 16 000/00
5000 0.300 0.370 0.328 11 11 11 35 35 23 65 65 30 000/00
6000 0.380 0.370 0.248 11 11 10 43 43 30 65 65 35 000/00
7000 0.350 0.270 0.479 9 9 8 33 33 25 120 120 60 000/00
8000 0.200 0.400 0.399 8 8 8 26 26 18 120 120 60 000/00
9000 0.231 0.301 0.467 13 13 13 132 132 105 60 60 30 000/00

182 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 USING GRADE UNCERTAINTY TO QUANTIFY RISK IN THE ULTIMATE PIT DESIGN

The North Hard Ore (zone 8000) predominantly underlies the
North Saprolite, and has the same structure, though not the
same orientation. This is most likely a result of insufficient
sample data to define a trend.
• All the hard sulfide variograms have isotropic horizontal
structures, attributable to the absence of close spaced grade
control data necessary to define the structures causing
anisotropy.
• The waste variograms for the saprolite (zone 2000) and hard
sulfide (zone 9000) are similar because of a relative lack of
grade variability between sample locations. This is to be
expected, considering that most of the mineralisation had
been included in the mineralised envelope.
Representative number of realisations
To determine the number of realisations that would be required to
obtain a reliable uncertainty model, the number of realisations
were plotted against the progressive mean and progressive
Coefficient of Variation (COV) for each simulated domain. The
number of simulated blocks for each domain was greater than
95 per cent of the total number of available blocks. For the first
few realisations, the mean and COV values in the plots ‘bounced
around’, but as the progressive values included more realisations,
they stabilised, and where they levelled sufficient realisations, they
were considered to have been used to model the variability. In this
study, the grade for five out of the nine simulation domains
could have been considered sufficiently simulated with up to
25 realisations, however, the remaining four domains required
additional realisations. Since it was easier to deal with
50 realisations than 25 realisations in the probability calculations,
and most of the Deep Sulfide ‘ore’ zones required more than
25 realisations, it was decided to use 50 realisations for all the
mineralised domains, and 25 realisations for the ‘waste’ domains.
Thus the 25 realisations for ZONECODE 2000 and 9000 were
repeated for the second 25 (26 to 50) mineralised domain
realisations. In this way, the simulation time was reduced by not
performing unnecessary simulations on the ‘waste’ domains, while
simultaneously achieving sufficient variability in the ‘ore’
domains, which were relatively quick to simulate.
Initial validations on point data
The internal nodes were estimated on 2.0 m × 2.0 m × 1.25 m
centres for each simulation, totalling 100 nodes per regularised
(SMU) block. These were only output for the first ten
simulations in order to check that the results honoured the
conditioning data, the histogram of the normal score and raw
data, and the variograms of the normal score and raw data. For
the remaining simulations, only the block values were retained.
Raw grade values were input into the program. The DBSIM
algorithm transformed them to normal score values for
simulation, and then back-transformed the resultant normal score
values to raw grade values. Both the raw and normal score data
sets were output, though because they show the same trends, only
the raw data is presented below.
Statistical validation
The base statistics for the raw grade data are summarised in
Table 4.
From Table 4, the base statistics indicated that the mean grades
and variances for the saprolite material (ZONECODES 1000
to 4000) were generally slightly under-estimated with the
conditioning data, exhibiting values slightly higher than the
simulated values. Though the mean grades for the Saprolite
waste (ZONECODE 2000) and North Saprolite ore
(ZONECODE 3000) were only marginally lower, both the mean
grade and variance for the South Saprolite ore (ZONECODE
4000) were significantly lower than the corresponding
conditional data values. This was initially a cause for concern for
the pit optimisation phase of the project because the South
Saprolite ore comprises the majority of the saprolite material
below the current LOM pit. However, the total saprolite ore only
comprises 16 per cent of the total ore tonnes between the current
LOM and Deep Sulfide pit designs. The far north Saprolite
(ZONECODE 1000) is relatively insignificant regarding its ore
tonnage contribution towards the Deep Sulfide Project.
From the base statistics, the simulation mean and variance for
the main ore high grade and low grade (ZONECODEs 5000 and
6000 respectively) – the primary drivers for the Deep Sulfide
Project – though marginally high, are reasonably close to the
mean and variance of the conditioning data, which fall within
the range of values of the ten realisations. Similarly for the
Hangingwall and Footwall mineralisation (ZONECODE 7000),
the conditioning data mean and variance are within acceptable
limits of the corresponding ten realisation values. The
simulated mean and variance values for the North Hard
ore (ZONECODE 8000) are significantly higher than the
corresponding conditioning data. A poor result, however, is to
be expected since this domain has been insufficiently sampled.
Furthermore, the North Hard ore comprises a relatively
insignificant proportion of the total Deep Sulfide ore and
therefore is not expected to have significant influence on
the pit optimisation runs. For the Hard Sulfide waste zone
(ZONECODE 9000), the simulated mean grade (0.50 g/t to
0.54 g/t) was significantly higher than the mean of the
conditioning data (0.26 g/t), though the variance is similar. This
zone was expected to remain largely subeconomic and was
therefore unlikely to impact significantly on the pit
optimisation runs.

TABLE 4
Summary base statistics for raw grade point data.

ZONECODE Conditioning data Simulation range in mean value Simulation range in variance
Mean Variance Minimum Maximum Minimum Maximum
1000 1.65 7.27 1.44 1.62 4.87 6.55
2000 0.36 1.69 0.31 0.33 0.72 1.21
3000 2.03 25.74 1.86 2.02 17.01 23.36
4000 2.93 32.18 2.62 2.71 19.99 25.13
5000 2.84 8.73 2.82 3.03 8.84 16.43
6000 1.36 6.26 1.35 1.44 5.68 7.27
7000 1.14 6.94 1.17 1.33 6.24 10.65
8000 2.09 28.49 2.16 2.54 19.30 47.66
9000 0.26 0.54 0.50 0.54 0.54 0.73

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 183
S ROBINS

Histogram validation
Histograms were generated for the first ten realisations of both
the raw grade and normal score simulation nodes. Since the two
data sets are similar, only the raw data is discussed. The raw
grade histograms for the saprolite material (ZONECODES 1000
to 4000) indicated that, except for the North Saprolite ore
(ZONECODE 3000), the simulated grade for the saprolite
material has been understated. The Saprolite material
(ZONECODES 1000 and 4000) was only marginally so, but the
grade for the Saprolite waste (ZONECODE 2000) has been
significantly understated. Since this is a ‘waste’ domain, with
only five per cent of the conditioning data above an economic
cut-off grade to begin with, this grade understatement is not
expected to have a significant impact on the pit optimisation
runs. For the primary drivers of the Deep Sulfide Project, the
Main Ore High Grade and Low Grade (ZONECODES 5000 and
6000 respectively), the raw grade histograms show good
correlation between the ten realisations and the conditioning
data. Figure 4 illustrates this deduction for the main high-grade
mineralisation (ZONECODE 5000).
Cumulative Frequency Distribution - ZONECODE 5000 Raw Grade Data
105%
100%
95%
90%
85%
Cum ulative Frequency (%)

80%
75%
70%
65% FIG 5 - Normal score variogram model for ZONECODE 5000.
60%
55%
50%
45%
40% model for the Main High-Grade mineralisation (ZONECODE
35% 5000). This was generally the case for the other zones,
30%
25% particularly the Hard Sulfide mineralised domains
20% Realisations 1 - 10
15% Conditioning Data
(ZONECODES 5000 to 8000).
10%
0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30
Upper Class Limit (g/t) Post simulation processes
After the simulation process, the base statistics per block for all
FIG 4 - Cumulative frequency distribution for ZONECODE 5000 the realisations were calculated. The following statistics were
raw grade data. calculated for each block:

For the Hard Sulfide Hangingwall and Footwall mineralisation

(ZONECODE 7000), the simulated grade appears to be mean, m =
∑ AU (1 to n)

marginally overstated between 3 g/t and 5 g/t. The same is n

evident for the North Hard ore (ZONECODE 8000) between
5 g/t and 7 g/t. Both of these domains require additional drilling where:
to ensure they are representatively sampled. The histograms for n is the number of realisations
the Hard Sulfide ‘waste’ domain (ZONECODE 9000) show that
the grade is significantly overstated up to about 2.5 g/t, however
94 per cent of the conditioning data is below the lower marginal variance, s2 =
∑ (AU − m) 2

cut-off grade of 1.20 g/t. The overstatement of grade is therefore (n − 1)

unlikely to have a significant impact on the pit optimisations.
From a statistical perspective, the Sadiola simulation model can standard deviation, s = s2
be considered to be representative of the conditioning data.
s
Variogram validation coefficient of variation, COV =
m
The final test to determine the validity of the simulation model
was to test the reproduction of the spatial correlation between the Similarly, the probability of each block grade being above 2.0 g/t
simulated results and the conditioning data. Experimental was calculated.
variograms were calculated for the first ten realisations and
plotted against the semi-variogram model that was input into the No of realisations > 2.0g/t
simulation process. To facilitate this process, the DBSIM probability, P(AU > 2.0 g/t) =
program converted the input semi-variogram models to block Total no of realisations (n)
support and output the results. The resultant normal score
variogram model for the Main High-Grade mineralisation (zone Next, for pit optimisation purposes, the laterite domains were
5000) is presented in Figure 5. added to the simulation model, with each realisation adopting the
Although not an exact match, the normal score simulated estimated grade for the laterite model blocks. The final simulation
experimental data compares well with the regularised variogram model was depleted by the saprolite life of mine volume.

184 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 USING GRADE UNCERTAINTY TO QUANTIFY RISK IN THE ULTIMATE PIT DESIGN

The final simulation model contained primary fields for
50 grade realisations, coefficient of variation (providing an
indication of uncertainty in the grade value), and probability that
the grade would be greater than 2.0 g/t. Secondary fields were
the remaining base statistics, the initial classification and rock
type fields, and the ZONECODE field. A representative section
(EW-5550) is presented to illustrate the primary fields of the
final simulation block model (Figure 6 to Figure 8). The $475/oz
prefeasibility pit shell is shown in white, and the models have
been depleted to the current ‘oxide’ LOM design.
From Figures 6 and 7 it is evident that although realisation one
shows significant material within the prefeasibility pit shell
above 2.0 g/t, the probability of the actual grade being above
2.0 g/t is low. For this reason, one shouldn’t place too much
emphasis on the grade of a single realisation, but rather consider
a number of realisations simultaneously. The use of probability
calculations provides a good overview of the results of all the
realisations simultaneously.
Figure 8 illustrates how the coefficient of variation calculation
provides an indication of uncertainty in the simulated block grade.
The diagonal white lines are borehole traces, and it is evident that
zones of high confidence (low uncertainty) correspond well with
the borehole traces, while zones of high uncertainty correspond
with zones that have been insufficiently sampled. Since the
coefficient of variation provides a reliable indication of uncertainty
in the simulated grade, it can also be used as a tool for confirming,
or even improving, the classification of the Deep Sulfide
mineralisation. Furthermore it has potential to direct drilling
programs in zones that require additional sample information, for
example, the Footwall zone to the east.
PIT OPTIMISATIONS AND DATA PROCESSING
For each simulated orebody model, an ultimate pit shell, or
maximum cash flow pit, was derived using the Lerch-Grossman
algorithm (Lerchs and Grossmann, 1965). The algorithm
essentially begins at the model surface and ‘mines down’ into the
model making two basic decisions for each block:
1. Should the block be mined?
2. Does the material go to the treatment plant or the waste
dump?
These two decisions are essentially based on the revenue
generated by the block at a defined gold price, which is offset
against the cost of mining the block and the cost of processing
the material. Although valued individually, the decision to mine a
block or not is determined by considering the surrounding blocks
in conjunction with the block in question. For example, mining a

W E

FIG 6 - Simulated block grade for realisation 1 on Section EW-5550.

W E

FIG 7 - Probability of grade greater than 2.0 g/t on Section EW-5550.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 185
S ROBINS

W E

FIG 8 - Block coefficient of variation on Section EW-5550.

high grade block below waste blocks may in fact pay for the cost
of mining the waste blocks and still generate a profit. The
ultimate pit is thus generated from all the blocks that the
optimiser decides to mine. The pit optimisation parameters were
used to optimise the simulation realisations (Tables 5 and 6).
The 50 NPV results for the various realisation ultimate pits
were plotted with the NPV for the recoverable resource model
(Figure 9). From Figure 9 it is evident that the NPVs for all the
simulation realisations are greater than the NPV for the
recoverable resource model and that most of them are
significantly greater. This indicates that significant upside
potential exists for the Deep Sulfide Project. Although this is
indicative of upside potential, it would be more useful to take this
information a step further and calculate probability or risk pits
from the individual realisation pits.
Calculation of probability/risk pits
The probability of mining each block was determined from the
number of times it was mined in each of the 50 ultimate pits.
Each block was then flagged in the simulation model, based on
its probability of being mined, and risk pit shells were generated
at five per cent probability intervals from five per cent to 95 per
cent (Figure 10 and Figure 11).
The $475/oz prefeasibility pit shell is shown as a white line in
Figures 10 and 11. It is evident in both figures that the
prefeasibility pit does not fetch a significant amount of high
probability/low risk material, and that in the east (Figure 10), it
mines low probability/high risk Footwall material. The Footwall
is known to be under-drilled and is the target for the Phase 8
Deep Sulfide drilling program.

TABLE 5
Pit optimisation parameters.

Gold price $475

Selling cost 7% (6% royalty + 1% management fee)
Annual throughput 5.2 Mt

Rock type Mining Processing

Mining unit cost Mining CAF Processing unit cost Recovery
US$/t US$/t
Laterite (ox) 1.70 1.02 17.05 0.93
Saprolite oxide – soft 1.67 1.00 16.30 0.93
Siliceous oxide – hard 1.58 0.94 22.67 0.93
Saprolite sulfide – soft 1.50 0.90 21.01 0.8
Hard sulfides – hard 1.32 0.79 20.29 0.8
Intermediate oxide 1.59 0.95 17.05 0.93
Intermediate sulfide 1.59 0.95 21.76 0.8
Mixed oxide 1.59 0.95 22.67 0.93
Mixed sulfide 1.59 0.95 21.52 0.75

$/tonne/bench (10 m)
Applicable to 160 level
Adjustments 0.0159

186 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 USING GRADE UNCERTAINTY TO QUANTIFY RISK IN THE ULTIMATE PIT DESIGN

TABLE 6 3.93 g/t), though the tonnes above cut-off show significant
upside potential.
Slope parameters for the pit optimisation.
• P80: The grade tonnage curve is similar to that of P50 at
North/south division Slope angle Domain Slope 2.0 g/t cut-off, with the SMU grade of 3.79 g/t falling in the
0 to 180 35 1 East oxides middle of the realisation grade distribution of 3.67 g/t to
180 to 359 40 1 West oxides 3.93 g/t. Upside tonnage potential is indicated above cut-off.
0 to 180 52 2 East sulfides • P95: Similarly indicates significant upside potential to the
tonnage above 2.0 g/t, with the SMU grade of 3.83 g/t
180 to 359 48 2 West sulfides
midway between the realisation grade limits of 3.74 g/t and
4.0 g/t.
Generation of grade tonnage curves Except for P05 – for which the average grade above 2.0 g/t
cut-off is marginally low - the above probability pit shells indicate
For each of the 50 realisations, and the recoverable resource that the SMU grade has been estimated with reasonable accuracy,
(SMU) model, grade tonnage curves were generated for the P05, even though within all the shells, the tonnes above cut-off have
P50, P80, P95, and prefeasibility probability pit shells. Less been understated in the SMU model. The tonnage and grade
variability than expected was achieved in each of the grade results are not too dissimilar despite the different techniques.
tonnage curves:
Within the prefeasibility pit shell, a comparison of the simulation
• P05: At the approximate economic cut-off grade of 2.0 g/t, and SMU model grade tonnage curves indicates that at 2.0 g/t
the SMU model grade was 3.70 g/t, while the simulated cut-off, both the SMU grade and tonnage curves lay within the
realisation grades varied between 3.65 g/t and 3.87 g/t. The range of the realisation distribution. This is illustrated in
realisations indicated significant upside potential for both Figure 12, where the tonnage curve has been re-scaled to a
grade and tonnage above cut-off. maximum of 70 million tonnes. Although one may argue that the
• P50: At 2.0 g/t cut-off, the SMU grade of 3.76 g/t falls in the SMU model is marginally too selective, Figure 12 indicates that
middle of the realisation grade distribution (3.65 g/t to this selectivity is acceptable. Above 2.0 g/t cut-off grade, the

NPV - Simulation Model Pits vs Recoverable Resource Model Pits

220
200
180
160
NPV (US$ Million)

140
120
100
80
60
40
20
-
1
3
5
7
9
11
13
15
17
19
21
23
25
27
29
31
33
35
37
39
41
43
45
47
49

Simulated Realisations Recoverable Resource Realisation Number

FIG 9 - NPV for simulation realisations and recoverable resource model.

FIG 10 - Plan view (elevation -90 m) showing the probability of mining.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 187
S ROBINS

W E

FIG 11 - Section EW-5550 showing probability of mining.

Grade Tonnage curves for the Prefeasibility Pit Shell TABLE 7

70 7
Ore/waste cut-off grades.
60 6
Average Grade Above Cut-

Rock type Cut-off (g/t) Rock type Cut-off (g/t)

50 5
Laterite 1.10 Hard sulfide 1.90
Tonnes (Millions)

off (g/t)

40 4
Saprolite 1.00 Intermediate oxide 1.10
30 3
Siliceous oxide 1.50 Intermediate sulfide 1.60
20 2
Saprolite sulfide 1.60
10 1

0 0 The calculated open pit values using the recoverable resource

0.0 0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.0 model for each of the probability pits (P05 to P95 at 0.5 intervals)
Cut-off (g/t)
are presented below in Figure 13. As expected, an inverse
SMU Tonnes SMU Grade
relationship exists between the pit value and stripping ratio with
increasing mining volume. There is an interesting kink in both
FIG 12 - Grade tonnage curves within the prefeasibility pit shell. curves for probability pit P80, which leads to its inclusion in
further pit value analyses. These additional analyses involved
comparing the values for each of the simulation model realisations
average grade above cut-off for the SMU model is 3.82 g/t, while with the SMU model value within selected probability pits. Open
the realisation grade ranges from 3.75 g/t to 3.99 g/t. At the same pit values were calculated for selected probability/risk pits (P05,
cut-off grade, the SMU tonnage of 19.5 Mt falls within the P50, P80, P95) and the $475/oz prefeasibility pit using all
realisation distribution range of 19.0 Mt to 21.0 Mt. 50 simulated realisations and the recoverable resource (SMU)
model. Table 9 summarises the pit value base statistics for the
Valuation of the pit shells 50 realisations and SMU model per pit shell, and Figure 14
illustrates the distribution of the realisation pit values.
A pit value was calculated for each of the probability pits (P05 to
P95 at 0.5 intervals) using the recoverable resource model. Since Probability Shell vs Open Pit Value
the extraction of the resource within these shells was not 50 10.6

scheduled, it was not possible to calculate an NPV for each pit, 40 10.4
30 10.2
Open Pit Value (US$) - (Millions)

but an open pit value was obtained for each scenario, based on 20 10.0
the cost of mining ore and waste, the cost of treating the different 10 9.8
Stripping Ratio

material types, and the revenue generated at a gold price of 0 9.6

$475/oz. The 2006 grade control cut-off grades were used to -10 9.4
-20 9.2
define ore and waste material for the various rock types
-30 9.0
(Table 7), with material above the cut-off categorised as ore. -40 8.8
The cut-off grades correspond with ‘break-even’ marginal -50 8.6
-60 Prob vs Val 8.4
grade material. The various parameters of recovery, cost, and Prob vs SR
-70 8.2
revenue that were used in the pit valuation calculations are -80 8.0
presented in Table 8. The cost values are those calculated by the 5 10 15 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95
Sadiola long-term mine planner in February 2006. The Probability Shell / Risk Pit

rehabilitation (rehab) costs had been included in the variable

process costs and were therefore set to zero. FIG 13 - Open pit values for the probability pits.

188 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 USING GRADE UNCERTAINTY TO QUANTIFY RISK IN THE ULTIMATE PIT DESIGN

TABLE 8 Histogram of Open Pit Profit Values

Cost parameters applied during pit valuation. 22

20
Item Material type Value
18
Mine call factor 100%
16
Recovery saprolite oxide 93%
14

Frequency
Recovery saprolite sulfide 80%
12
Recovery hard sulfide 80%
10
Fixed mining ($/tonne treated) 2.40
8
Fixed process costs ($/tonne) 1.07 6
Laterite (oxide) 9.84 4
Saprolite oxide – soft 9.09 2
Siliceous oxide – hard 15.46 0
Variable process costs ($/tonne) Saprolite sulfide – soft 13.80 -20 0 20 40 60 80 100 120 140 160 180

Hard sulfide – hard 15.58 Class Mid Point - Profit ($ Million)

Intermediate oxide 9.84 P05 P50 P80 P95 PPF

Intermediate sulfide 14.56

FIG 14 - Histogram of the open pit values for the 50 realisations.
Admin costs ($/tonne) 3.59
Capital replacement ($/tonne) 0.46
Pit Values for Simulations and SMU Model
Rehab costs ($/tonne) 0.00 200
Laterite (oxide) 2.089
150
Saprolite oxide – soft 2.058 Pit Value ($ Millions)
100
Siliceous oxide – hard 1.938
Variable ore ($/tonne) Saprolite sulfide – soft 1.852 50

Hard sulfide – hard 1.825 0

Intermediate oxide 1.960 -50

Intermediate sulfide 1.960 -100

P05 P50 P80 P95 PPF
Laterite (oxide) 2.083
Probability Pit
Saprolite oxide – soft 2.051 Realisation (1 to 50) SMU Model
Siliceous oxide – hard 1.933
Variable Waste ($/tonne) Saprolite sulfide – soft 1.846 FIG 15 - Pit value per probability pit for the simulation and SMU
models.
Hard sulfide – hard 1.820
Intermediate oxide 1.954
confirms that the prefeasibility pit is the optimal pit for the Deep
Intermediate sulfide 1.954 Sulfide project, given the $475/oz gold price and cost
Gold price ($/oz) 475 parameters. The SMU model produces a profit of $112 million
and falls in the middle of the distribution of the simulated
realisations. The risk does exist that a profit of only $64 million
TABLE 9 may be realised, however a profit of $179 million is also
Base statistics for the realisation pit values per pit shell. possible, with the most likely profit ranging between $87 million
and $148 million. It is evident from Figure 15 that the profit
Pit SMU Mean Std dev Max profit Min profit curves flatten off significantly from the P50 to the P95 risk pit,
profit profit ($ million) ($ million) ($ million) and that there is no insignificant difference in their mean profit
($ million) ($ million) (Table 9). Though a marginal decision at a gold price of $475/oz,
P05 -78.010 -2.621 34.782 86.379 -67.288 mining an equivalent of the P80 or P50 pits could dramatically
P50 14.011 85.385 31.119 160.421 30.655
improve the Deep Sulfide reserve ounces, and the P50 pit would
most probably become a real option should the gold price
P80 21.745 82.204 24.921 139.124 37.691 increase significantly.
P95 42.085 89.025 21.389 136.528 47.883
P-PF 112.889 108.036 24.400 179.400 60.870 CONCLUSION
The final risk pits generated from the simulation model indicated
From Figure 14 and Table 9 it is evident that the $475/oz that the prefeasibility pit is generally conservative, corresponding
prefeasibility pit offers the greatest profit margin but not the least with the 100 per cent probability shell. There is however a
risk. Probability pit P95, not surprisingly, has the tightest certain amount of risk involved with mining the Footwall
distribution about the mean and would therefore offer the least mineralisation, where the prefeasibility pit digs on
risk to achieving the indicated profit. However, considering the high-risk/low-probability material. This is confirmed by the COV
respective mean profit values for the prefeasibility and P95 pits, values in this zone, which indicate high uncertainty in the grade
the value offered by P95 is significantly lower than that of the estimate. Based on these results, an infill drilling program has
prefeasibility pit. Given the relatively minor additional risk, the been planned to mitigate this risk. At the current economic
prefeasibility pit may be the preferred pit option. Figure 15 cut-off grade of 2.0 g/t, grade tonnage curves generated for

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 189
S ROBINS
selected risk pits have shown that although upside tonnage
potential exists for the Deep Sulfide Project, the recoverable
resource (SMU) model grade estimate is reliable. With the
current cost information, and a gold price of $475/oz, the
prefeasibility pit is optimal. The calculation of economic pit
values for the realisations within the same selected risk pits
indicated that the P50 to P95 pits and the prefeasibility pit are
not significantly different. Therefore, there is potential to add
reserve ounces to the project without significantly decreasing its
value. An increase in gold price, or a decrease in operating costs,
would make this additional material available to the project. At a
gold price of $475/oz, and with current cost parameters, the
prefeasibility pit shows maximum value for the project with no
significant additional risk.
This study has highlighted the potential to use the COV values
as a reliable classification method. It is recommended that the
classification is re-run using the simulation COV information, as
proposed by Dohm (2005). This exercise would eliminate
unnecessary drilling and potentially save drilling costs. The gold
price used for this project was $475/oz (the recommended
Anglogold Ashanti resource reporting gold price for 2005).
Anglogold Ashanti has since increased their resource reporting
gold price to $1000/oz, and at $720/oz, their reserve reporting
gold price is significantly greater than the gold price used for this
project. Since the current gold price provides significant upside
potential for the Deep Sulfide Project, it is recommended that the
pit shell work is rerun using a suggested gold price of $700/oz
and the latest mining and processing costs derived from the
prefeasibility project.
190 Spectrum Series Volume 17
REFERENCES
Benndorf, J and Dimitrakopoulos, R, 2007. New efficient methods for
conditional simulation of large orebodies, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 103-110 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Boucher, A and Dimitrakopoulos, R, 2009. Block-support simulation of
multiple correlated variables, Mathematical Geosciences,
41(2):215-237.
Dimitrakopoulos, R, Farelly, C T and Godoy, M, 2002. Moving forward
from traditional optimisation: Grade uncertainty and risk effects in
open-pit design, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 111:A82-A88.
Dohm, C, 2005. Mineral resource classification from innocence to
excellence and beyond, presentation made to the Anglo American
Mine Geology Conference 2005, MinRED, 24 p.
Godoy, M, 2003. The effective management of geological risk in
long-term production scheduling of open pit mines, PhD thesis
(unpublished), University of Queensland, Brisbane.
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
pp 370-382 (Oxford University Press: New York).
Lerchs, H and Grossmann, L, 1965. Optimum Design of Open-Pit Mines,
Trans CIM, LXVII:17-24.
Peattie, R, 2005. DBSIM – Direct Block Simulation Algorithm User’s
Manual (Draft 1), unpublished document, AngloGold Ashanti, South
Africa, 18 p.
Robins, S P, Nicholls, C E, Fouche, F and De Hert, G, 2005. Sadiola –
Identified Mineral Resource Statement 2005 – Part Two A – Sadiola
Main Pit and Hard Sulfides, unpublished company report, SEMOS,
SA, Mali, pp 38-39.
Van der Westhuizen, R, 2005. Sadiola Reserve Report December 2005,
unpublished company report, SEMOS, SA, Mali.
Advances in Orebody Modelling and Strategic Mine Planning I

Angle Reverse Circulation Versus Blastholes in Ore Control –
Sampling Quality Versus Data Density – A Case Study
I Douglas1, B Perry2, W Hardtke2 and L Allen3
ABSTRACT
Many open pit mines face a difficult choice between the cost of special
drilling programs for better sampling of ore control and sampling the
production drilling, usually with blastholes. This choice is not always
made on the basis of detailed studies, but on past experience and any
biases that those charged with making the decision may have. This can
lead to significant costs, either in the execution of unnecessary drilling
programs or poor definition of ore and waste – sometimes both. This
paper outlines a methodology used by Newmont Mining to complete a
cost-benefit study and present results of the study from a theoretical
viewpoint, as well as from production. The approach included generating
conditional simulations, extraction of various vertical blasthole and angle
reverse circulation (RC) data sets representing potential configurations,
completing ore control utilising different methodologies and tabulation
and comparison of results. The study concluded that blasthole sampling
was the preferred approach for this deposit. The conclusion was a
function of:
• a planned blasthole spacing that is quite close,
• the close spaced data allowing the information of surrounding
blasthole assays to correct for local sampling variances coming from
poorer quality sampling, and
• the prohibitive cost of close spaced angle RC spacing needed to
correctly classify ore and waste.
It is acknowledged that this would likely not be the conclusion if
blasthole sampling were significantly biased. Actual production data has
demonstrated that systematic biases are not present in the blasthole
samples. Mill reconciliation to production demonstrates that the polygons
are well defined, and excessive dilution from misclassification is not
occurring.
INTRODUCTION
Newmont acquired the Ahafo project through its merger with
Normandy Mining in 2002. Following a successful feasibility
study in 2003 and construction in 2004 - 2006, Newmont
commenced mining operations from two open pits, Apensu and
Subika, in Ahafo in January 2006, with plant wet commissioning
in June 2006. During 2007, mining products consisted of 8 Mt of
oxide, transitional, and primary ore, grading 1.88 g/t Au and an
associated 38 Mt of waste. In 2005, finalisation of the ore control
system and approach were under review. The mining approach
selected was to drill and blast on 9 m benches and mine on 3 m
flitches. Two options were considered for ore control approach in
primary ores:
1. drill three 9 m benches with angle reverse circulation (RC)
sampling on 2 m intervals, or
2. sample the blastholes (9 m) and split into three 3 m samples
for assaying.
Because no blasthole drilling was completed in oxidised
saprolite ores which were free-digging, a full set of angle RC
data was completed in oxidised rock, which formed the
conditioning data for simulation studies completed at Apensu.
1. MAusIMM, Newmont Mining Corporation, 10101 E Dry Creek Road,
Englewood CO 80112, USA. Email: ian.douglas@newmont.com
2. Newmont Mining Corporation, 10101 E Dry Creek Road, Englewood
CO 80112, USA.
3. MAusIMM, Newmont Mining Corporation, 10101 E Dry Creek Road,
Englewood CO 80112, USA. Email: larry.allen@newmont.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
Other issues were of concern in choosing the best approach to
ore control, such as obtaining good samples from blastholes and
the possible presence of water in the mining areas. These studies
were completed to help guide the decision regarding which
approach to take in primary ores. The ore zones were noted to be
continuous between saprolite and primary rock. Initial
reconciliation in the oxidised ores was poor, with ore control
finding more tonnes at lower grade than predicted by the
exploration model. Blasthole sampling was introduced after nine
months of continued poor reconciliation within primary rock,
with the site making a full change to blasthole ore control in
March 2007. Improvement in reconciliation was measurable after
three months.
LOCATION AND GEOLOGIC SETTING
The Ahafo gold property is located in western Ghana, between
the towns of Kenyase and Adroboa in the Brong-Ahafo Region,
about 290 km northwest of Accra. It is 107 km northwest of
Kumasi, and 40 km south of the regional capital of Sunyani
(Figure 1). The project is composed of a southern and northern
area, with the plant site and current operations located in the
southern area. The Apensu deposit that is the subject of the
conditional simulation study (Figure 2) is also located in the
southern area, adjacent to the plant site. The Ahafo project is
located in the West African Craton, within the Palaeoproterozoic
Man Shield. The project is located within the Shield, in the
eastern terrain underlain by Palaeoproterozoic Birimian rocks.
The Birimian rocks consist of five evenly-spaced tholeiitic to
acidic composition volcanic belts trending northeast-southwest.
Basins between the volcanic belts are filled by predominantly
turbiditic sediments and the transition zones between the
volcanic rocks and the sedimentary rocks are filled with
chemical sediments. Three granite successions have intruded the
Birimian rocks. The Ahafo project is located in the Sefwi Belt,
one of the five Birimian volcanic belts. The Sefwi Belt is
bounded to the northwest by the Sunyani Basin, and to the
southeast by the Kumasi Basin. Dixcove-type or ‘belt-type’
granitoids intrude the contact and predominate in the
metavolcanics that form more or less elongate bodies parallel to
the regional strike. Volcanic rocks in the belt are mainly basaltic
and have been metamorphosed to varying degrees from lower
greenschist to lower amphibolite facies. The sedimentary
succession consists mainly of fine to medium-grained lithologies
(argillites and wackes) with variable amounts of volcaniclastic
material. Cape Coast type granitoids intrude the metasedimentary
rocks. The strong north-easterly grain of the Palaeoproterozoic
terrane parallel to the volcanic belts is the predominant trend that
governs major structural features. The faults and associated
structures display a complex history of movement, including thrust
faulting and shearing with both normal and strike-slip motion and
have played a major role in emplacement of mesothermal gold
mineralisation. Regional structure is controlled by the Kenyase
Thrust Fault, a northeast to southwest trending regional thrust that
separates the Sefwi Belt from the Sunyani Basin.
Three discrete geological zones are recognised within the
Ahafo project, relating to currently identified mineralisation
styles. The Ahafo district currently consists of 12 discrete
deposits, localised along multiple northeast-striking structural
Spectrum Series Volume 17 193
I DOUGLAS et al
FIG 1 - Project location plan.
zones. All of the ‘shear zone’ type deposits appear to be part of
the same mineralised system. As with many deposits located in
tropical climates, a saprolite zone, typically between five and
50 m thick, is developed at the surface. The saprolite zone gives
way at depth to a sulfide zone, where gold occurs in structurally
controlled zones of hydrothermal alteration. Six lithologies and/or
litho-structural units have been recognised in the Kenyase-style
zones, consisting of:
1. weakly foliated mixed (meta)-pelitic sedimentary rocks and
(meta)-mafic volcanic units in the footwall of the Kenyase
Thrust Fault;
2. Dixcove Suite granitoids form the hanging wall to the
thrust;
194 Spectrum Series Volume 17
3. mylonites after granitoid of weak to strong foliation and
local folding within the Kenyase Thrust Fault;
4. multiple thrust fault duplexes consisting of mixed mylonite
after granitoid and volcano-sedimentary units, containing
rigid granitoid fragments in a matrix of plastically-deformed,
predominantly chloritic and volcano-sedimentary units;
5. phyllonite units, dominated by phyllosilicate minerals,
resulting from plastic deformation of the volcano-
sedimentary units; and
6. tectonic breccias (cataclasite) and finely-crushed rock
formed within the granitoid units by local, brittle reactivation
on or near the Kenyase Thrust Fault.
Advances in Orebody Modelling and Strategic Mine Planning I
 ANGLE REVERSE CIRCULATION VERSUS BLASTHOLES IN ORE CONTROL – SAMPLING QUALITY VERSUS DATA DENSITY

FIG 2 - General layout, Ahafo project.

Hydrothermal alteration in Kenyase Style deposits is
characterised by the alteration of primary chlorite to sericite-silica,
and the addition of silica, iron carbonate, pyrite and local albite to
the host rock. Ore zones, while structurally controlled, do not have
visual identifiers that can be used reliably to determine ore and
waste. Discontinuities are abundant at the ore control scale,
whereas overall geometry is reasonably continuous. An example
cross section of the Apensu deposit is depicted in Figure 3.
CONDITIONAL SIMULATION FRAMEWORK
In preparation for initial mining in January 2006, an angle RC
data set was available for use in generating conditional
simulations for determining future ore control options. Initial
review of the angle RC data set showed that while the main zone
was continuous, the ore-waste boundary, at approximately
0.6 g/t, was irregular, somewhat irregular on section and not very
predictable from section to section. The only available tool to
provide predictive capability for making decisions on how to
approach ore control was considered to be conditional
simulation. An area of approximately 300 m (along strike) by
150 m was drilled on approximately a 12.5 m × 12.5 m grid.
Holes were angled at 60° into the Kenyase thrust, approximately
perpendicular to the mineralisation (Figures 4a and 4b).
A set of ten simulations using the sequential Gaussian
simulation method was completed over the test area on 1 m × 1 m
horizontal node spacing and 3 m vertical spacing. The program
SGSim of GSLIB (Deutsch and Journel, 1998) was used. After
reviewing statistics and comparing them with the original data, a
single simulation was selected as the ‘ground truth’ based on the
closest visual and statistical comparison to the original RC data.
Once the base simulation was selected, vertical blasthole and

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 195
I DOUGLAS et al
FIG 3 - Cross-section through the Apensu deposit.
FIG 4 - Source data for conditional simulation. (A) Oblique view of
Angle RC and raw assays. (B) Cross-section displaying 3 m
composited grades and outline at 0.6 g/t.
angle RC data sets were extracted from the simulation (Figures 5a,
5b and 5c) and used in subsequent testing of ore control. The
blasthole data was extracted at a 6 m × 6 m spacing, and the
angle RC on a 12.5 × 12.5 m spacing. This allowed 36 different
blasthole data sets by moving the origin 1 m each time and 25
different angle RC data sets by moving 2.5 m each time.
196 Spectrum Series Volume 17
FIG 5 - Extracted data sets and source simulation. (A) Blasthole
data set, (B) source simulation and (C) angle RC data set.
Recognising that sampling of blastholes has numerous
challenges that usually result in less than an optimal sample being
taken, a random error component was tested on the blasthole data
set before its use in the ore control tests. A random error
component was added to the blasthole data set. This random error
was tested in ranges from ±10 to ±90 per cent. Errors from the ±10
to the ±50 per cent tests showed little difference in results, and at
±90 per cent, the difference was approximately two per cent on
contained ounces inside ore control polygons using the affected
data set. For the purposes of the study, a subjective choice of
±20 per cent random error was applied to the extracted blasthole
data set in all consequent tests. Ore control comparing the
blasthole data sets to the angle RC data sets was completed both in
an automated way as well as using a manual definition.
Advances in Orebody Modelling and Strategic Mine Planning I
 ANGLE REVERSE CIRCULATION VERSUS BLASTHOLES IN ORE CONTROL – SAMPLING QUALITY VERSUS DATA DENSITY
Automated ore control test
The automated approach removed the subjectivity of defining
polygons by hand and assumed that the polygons defined from
SMU block outlines would be mineable. This approach was
applied to both drill types (blasthole and angle RC). These
polygons were not directly compared to hand drawn polygons.
Both the blasthole data based ore control polygons and the angle
RC based ore control polygons were generated by ordinary kriging
to an SMU-sized block of 5 × 5 m horizontal and 3 m vertical. For
the blasthole set, 36 sets of polygons were generated from the 36
possible blasthole data sets. For the angle RC set, 25 sets of
polygons were generated from the 25 possible angle RC data sets.
The block outlines at the cut-off grade for all the possible data sets
were then used as the defined ore polygons (Figure 6).
FIG 6 - Kriged SMU blocks (A) and ore polygons created at 0.6 g/t
outline (B) from one possible data set.
Manual ore control test
The manual approach is far more time consuming to complete
than the automated polygons, thus it was completed on only one
each from the 36 blasthole data sets and 25 angle RC data sets.
Each data set used was chosen randomly. The data sets were used
to interpolate a 2 m × 2 m block model using ordinary kriging,
then polygons were manually to separate ore and waste at 0.6 g/t
(Figure 7).
FIG 7 - Manually drawn polygons. (A) Blastholes and polygons
and (B) angle RC and polygons.
ORE CONTROL RESULTS
The results from the two approaches reached the same overall
conclusion that the use of blasthole-based ore control would
perform better, as well as insights into questions frequently
confronted in setting up ore control procedures.
Advances in Orebody Modelling and Strategic Mine Planning I
Automated results
The polygon outlines derived from the SMU blocks at 0.6 g/t
were used to filter the conditional simulation into nodes routed as
ore and waste. By tabulating the nodes inside the polygons
(routed as ore), one could compare the results expected from the
two approaches with the ‘actual’ results. These were tabulated
for all 36 possible outcomes from the blasthole and 25 possible
outcomes from the angle RC data. The tabulations were
completed on a net revenue basis, with process cost, recovery
and drilling cost included in the revenues. Because the blastholes
will be drilled irrespective of the chosen method, the only cost
associated with the blasthole approach is the assaying cost.
However, the angle RC approach must be able to offset the cost
of drilling the holes since they will only be drilled if needed for
ore control.
As shown in Figure 8, the blasthole generated polygons
consistently outperformed the angle RC polygons. Further
testing was completed to determine the optimal spacing for the
angle RC holes if that approach was to be applied. For the angle
RC data set, tests were also run at 9 m × 9 m spacing, 5 m × 10 m
spacing, 6 m × 6 m spacing and 4 m × 4 m spacing. Figure 8
presents results for all the angle RC tests, as well as estimated
results (linearly interpolated) for other configurations as
compared to the 6 m × 6 m blastholes. As shown in the graph, the
gross revenue increases as the spacing is reduced, reflecting the
benefit of the added information (drill holes). For the angle RC,
offsetting the benefit is the cost of the drilling, resulting in an
optimal spacing of 5 m × 10 m. Both the 6 m × 6 m and
7 m × 7 m spacings appear to give similar results on a net
revenue basis. Due to the cost savings from using blastholes, the
net revenue from the blasthole polygons exceeds all outcomes
from the angle RC data sets. On a gross revenue, the 6 m × 6 m
blasthole data is approximately equivalent to the 5 m × 10 m and
7 m × 7 m angle RC data sets.
Revenue from SIM model
19.0
18.5
18.0
Revenue x 1,000,000
17.5
17.0
16.5
16.0
Angle(6)
blast(25)-20%
15.5 Angle(25)
Angled(25) 9m
15.0
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35
Model
FIG 8 - Tabulation of results from blasthole and angle RC data sets.
Manual results
As with the automated approach, the polygons defined on the
blasthole and angle RC data sets were used to tabulate simulated
nodes within the respective polygons and compare results. The
RC polygons predicted a higher revenue than that predicted by
the blasthole polygons by about three per cent. However, the
angle RC approach resulted in polygons with more tonnes at
lower grade as compared to the blasthole data. In addition, the
angle RC over-predicted grade as compared to the underlying
simulated nodes by five per cent, whereas the blasthole predicted
grade was overestimated by two per cent. This resulted in the
blasthole approach having a net revenue of approximately
five per cent more.
Spectrum Series Volume 17 197
I DOUGLAS et al
TABLE 1
Initial reconciliation results.
F1 – OC/exploration
Tonnes Grade (g/t) Ounces Tonnes
Oct 2006 - Mar 2007 101% 90% 91% 102%
Apr 2007 - Jun 2007 96% 91% 87% 97%
Initial ore control approach
Due to the uncertainty of sampling blastholes in wet zones and
properly sampling coarse gold especially in blastholes, the initial
decision was to adopt the angle RC approach after completion of
mining in saprolite. After sufficient production information was
obtained, the ore control approach was revised.
PRODUCTION RECONCILIATION RESULTS
As predicted by the simulation results, the initial results from the
applied ore control system resulted in more tonnes at lower grade
than expected. Further studies were required to attempt to pinpoint
whether the poor reconciliation was due to poor performance of
the long term model, due to poor ore control application, or both.
Initially, the ore control system was completed on blocks
approximately the same size as the SMU (5 m × 5 m blocks),
which was thought to result in high dilution as well as poor
routing. Improvements to the ore control system included using
smaller blocks in the underlying ore control model, and the
application of probability shells to reduce smoothing. However,
the grade tonnage distribution did not improve markedly, and the
conclusion drawn was that it was unlikely that 12.5 m spaced RC
data could reliably predict ore and waste at the 3 m selectivity
being mined. After approximately nine months of mining
production, the blasthole versus angle RC study was reviewed, and
a program was proposed to test whether or not blasthole sampling
would improve reconciliation. Blasthole (BH) sampling was tested
and a methodology developed for sampling. Comparisons to
existing RC data showed that BH data was defining ore and waste
zones missed by RC drilling, and that the assays were unbiased
when paired with closest angle RC assays. An underlying
assumption is that blasthole samples are unlikely to indicate ore
where no ore exists and vice versa, even if the sampling of the
blastholes was less than optimal.
Sufficient blasthole assays were acquired by January 2007, and
a decision was made to move to full blasthole sampling. The ore
control system was changed to use blasthole assays in March
2007. Within three months, a measurable change in reconciliation
was documented, as shown in Table 1. Reconciliation was tracked
in three ways:
1. F1 comparing ore control to exploration model,
2. F2 comparing mill to mine, and
3. F3 comparing mill to exploration.
While tonnage defined in ore control did decrease by
approximately five per cent as predicted when comparing the
exploration model to ore control, there was not an associated
increase in grade, with the grade increasing only one per cent,
resulting in an apparent decrease in contained ounces by
approximately four per cent (F1, Table 1). However, when the
reconciliation of the mine to the mill is viewed, there was a seven
per cent increase in grade delivered to the mill as compared to that
predicted by ore control (F2, Table 1). Thus the expected increase
in grade not realised in ore control was in fact seen in the mill,
leading to the conclusion that the ore control was underestimating
grade. When combined, and comparing the mill back to the
exploration model, (F3, Table 1), there is a clear decrease in
tonnage and an increase in grade. While three months is not a
sufficient time frame to have high confidence, it was an indication
that the ore control change was having the desired effect. While
198 Spectrum Series Volume 17
F2 – Mill/OC F3 – Mill/exploration
Grade (g/t) Ounces Tonnes Grade (g/t) Ounces
102% 104% 102% 92% 95%
109% 106% 93% 99% 93%
the mill has continued to report higher grade than defined by the
ore control system (approximately six per cent higher in the
12 months to August 2008), the overall reconciliation to the
exploration model has been showing the model continues to
under-predict tonnages and over-predict grades, but resulting in
approximately five per cent higher contained metal over the
12 month period.
OTHER RESULTS
The automated ore control tests also investigated the impact of the
number of samples used in ore control. When originally
comparing the simulation re-blocked to SMU size blocks, the
kriging plan suggested that to achieve the correct grade tonnage
distribution, a restricted number of samples (six) should be applied
in the ore control, or risk over-smoothing the results. Tests were
run for both the blasthole and the RC data, comparing results
between using six, 15 and 25 composites. When comparing the
resulting outcomes of the simulated nodes within the polygons, the
use of 25 composites was superior by approximately 0.5 per cent
on revenue as compared to 15 composites and up to three per cent
as compared to using six composites. The use of grade envelopes
on the angle RC data set was also investigated in the manual test.
Grade envelopes were drawn in section and plan on the angle RC
data set prior to kriging in ore control at 0.6 g/t and 0.4 g/t.
Polygons were then defined on the resulting kriged estimates.
While the 0.6 envelope approach did define higher grade material,
the envelope caused the grade estimate to be severely overstated
when compared to the simulated nodes by approximately 20 per
cent, and to route excessive ore tonnage to waste, resulting in
significantly less revenue. When compared to completing ore
control without predefined envelopes, the results from no envelope
were superior to the 0.4 g/t envelope approach by two per cent on
revenue, and seven per cent better than using the 0.6 g/t envelope.
In both cases, the reason for the superior results was routing less
ore to waste.
CONCLUSIONS
The use of conditional simulation was successful in predicting the
outcomes from alternate ore control approaches. This has been
evidenced by the following. Firstly, field tests and Mine to Mill
reconciliation results. Secondly, blasthole-based ore control has
delivered to mill fewer tonnes at a higher grade – as validated
through reconciliation. Thirdly, errors in blasthole assaying are
being minimised through appropriate kriging plans in ore control
though the usage of sufficient samples to overcome the inherent
randomness in blasthole sampling; and lastly, measurable cost
savings have occurred by eliminating the angle RC drilling cost
and reducing the tonnage to be processed. While in this study the
blasthole approach is favoured over the angle RC approach, it is
acknowledged that results could be very different if ore waste
boundaries were more continuous, predictable and visually
mappable, as is seen in many shear zone type gold deposits.
Newmont will continue to review its ore control approach by
closely monitoring Mine to Mill reconciliation to improve results.
REFERENCES
Deutsch, C V and Journel, A G, 1998. GSLIB, Geostatistical Software
Library and User’s Guide, second edition, 369 p (Oxford University
Press: New York).
Advances in Orebody Modelling and Strategic Mine Planning I
 HOME

Impact of Bench Height, Cut-Off Grade and Production Spacing

on Misclassification in Ore Control – Challenges Produced by
High Nugget, Low Continuity Ores
W Hardtke1, L Allen2, B Perry3 and I Douglas4

ABSTRACT
Model reconciliation problems at the low-grade, high nugget Phoenix
gold-copper deposit led to a thorough examination of the influence of
production decisions on ore waste classification. A conditional
simulation study of the deposit was used as the basis for: quantification
of dilution at different bench heights, quantification of ore waste
misclassification relative to bench heights and cut-off grades, and
determination of an appropriate number of samples to optimise ore
control polygon definition.
Actual reconciliation data were empirically used to support
conclusions. This paper presents the results of the study and analyses the
relationship between misclassification of ore waste blocks and cut-off
grade. The paper also explores the implications on long-term mine
planning.

INTRODUCTION
The Phoenix deposit is a gold-copper deposit located in Northern
Nevada, USA. Gold mineralisation is characterised as low-grade,
but with a significant coarse gold component controlled by
localised high-angle structures within broad, shallow-dipping
sedimentary formational controls. The inherent variability of this
deposit has resulted in particular challenges in regards to
reconciliation between the resource model, ore control and the
mill. This is reflected by high nugget effects, low continuity in
the variography and a tendency to overestimate the gold grade in
the resource model.
Operations have recently changed from 6 m to 12 m bench
height to improve rock fragmentation, which has resulted in
improved mill throughput. However, with the 12 m bench height,
attempts to raise the head grade by increasing the cut-off did not
result in a higher grade at the mill. Conditional simulation was
used as a tool to quantify the difference in dilution and
misclassification caused by the larger bench height. It was also
used to optimise ore control parameters such as the number of
blasthole samples and the use of cap grades.
LOCATION AND GEOLOGICAL SETTING
The Phoenix deposit is located near Battle Mountain, Nevada,
USA (Figure 1). The deposit area was previously mined for
copper and gold by Duval Corporation between 1964 and 1982
and by Battle Mountain Gold for gold beginning in 1984. The
deposit came under Newmont control through the 2002 merger
between Newmont and Battle Mountain Gold. Mineralisation
1. Senior Consulting Geostatistician, Newmont Mining Corporation,
10101 E Dry Creek Road, Englewood CO 80112, USA.
Email: william.hardtke@newmont.com
2. MAusIMM, Director of Reserves, Newmont Mining Corporation,
10101 E Dry Creek Road, Englewood CO 80112, USA.
Email: larry.allen@newmont.com
3. Senior Consulting Geostatistician, Newmont Mining Corporation,
10101 E Dry Creek Road, Englewood CO 80112, USA.
Email: bruce.perry@newmont.com
4. Group Executive – Reserves and Geostatistics, Newmont Mining
Corporation, 10101 E Dry Creek Road, Englewood CO 80112, USA.
Email: ian.douglas@newmont.com
FIG 1 - Location map (Newmont Nevada Properties).
occurs within the five kilometres long, 60 to 70 W-dipping,
north-south Copper Canyon structural corridor with cumulative
apparent normal displacement of 650 to 825 m as described by
Newmont in the annual Phoenix reserve report (2006). This 250
to 550 m wide corridor is bound on the west by the Canyon
fault, with 600 m of apparent normal offset, and to the east by
the ore-controlling Virgin fault, with 100 to 250 m of apparent
normal offset. Stratiform skarn mineralisation is cut by
high-angle west-dipping faults. Au, Cu and Ag reserves are
contained mostly in the Pennsylvanian-Permian Antler overlap
sequence, which generally consists of a fining upward sequence
of ferruginous conglomerate of the Battle Formation, limestone
of the Antler Peak Formation and calcareous sandstone of the
Edna Mountain Formation. The Cambrian Harmony Formation,
consisting of sandstone and siltstone, is unconformably overlain
by the Antler sequence rocks and is only weakly mineralised.
Allochthonous argillite and chert of the Permian Havallah
sequence overlie the Antler sequence rocks in the upper plate of
the Golconda thrust.
Gold-copper reserves at Copper Canyon are roughly divided
into two by the Copper Canyon stock. Pyrrhotite-pyrite skarn of
the Antler Peak Limestone hosts high-grade gold ore in the
Fortitude orebody north of the stock, whereas lower gold grades
are present in the underlying Battle Formation. In contrast,
Au-Cu-Ag mineralisation in the Bonanza area south of the
stock is hosted mainly in skarn of the Battle Formation and
particularly in the lowermost conglomerate unit. Only a small
amount of Antler Peak Limestone is recognised south of the
stock and no Edna Mountain Formation is present.
Fracture-controlled Au-Cu mineralisation occurs along the
north, south and west margins of the Copper Canyon stock in
hornfelsed Havallah sequence in the West Bonanza, Reona and
Sunshine deposits, respectively.
CONDITIONAL SIMULATION FRAMEWORK
Conditional simulation studies have traditionally been used to
evaluate uncertainty as discussed by Goovaerts (1997). However,

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 199
W HARDTKE et al

the use of a single simulation is useful to provide a ‘ground-truth’

representation of a deposit. In order to investigate ore-waste
classification and misclassification issues, a conditional
simulation was created to represent the F1 area at Phoenix. This
area was selected because of the quantity of production data
available and because it had relatively high gold and low copper
grades. Blasthole data from 6 m benches were used to condition
the simulation, with only six and 12 m bench heights tested. The
4.8 m spaced blasthole data provided the necessary resolution
essential for examining misclassification at a detailed scale.
Several conditional simulations were developed for both gold
and copper within the F1 deposit using the sequential Gaussian
method in the GSLIB program SGSIM as presented by Deutsch
and Journel (1998). The variogram models and the search criteria
were developed using the Isaaks & Co Sage program (2001).
Nodes were simulated on a 1.2 m × 1.2 m horizontal spacing and
a 6.0 m vertical spacing. A single simulation was then selected to
represent the ground truth for the F1 deposit. This simulation
was selected because it not only had the same mean and variance
as the blasthole data but was visually most representative. The
visual validation was very important because the high nugget
effect tended to minimise the continuity of the mineralisation. A
typical bench from the simulation is shown in Figure 2.

ANALYSIS OF PRODUCTION DECISIONS USING

SIMULATED NODES
The Phoenix deposit has some difficult mining issues because it
is low-grade with coarse gold in ore zones that have only limited
continuity at the cut-off grades required for the flotation mill.
The combination of the inherent variability of the deposit,
sampling problems due to coarse gold and a relatively high
cut-off grade have led to excessive misclassification of ore and
waste. Possible solutions for these issues would include cost
reduction measures to lower the cut-off grade or increasing
FIG 2 - Plan view of selected simulation.
efficiencies in ore control by reducing the misclassification. With
the use of conditional simulations it is possible to quantify the
dilution associated with larger bench heights and as well as the cut-off grade and affects block revenues, copper is generally
misclassification due to the intrinsic problems caused by the ignored in this paper as the degree of ore waste misclassification
variability of the deposit. results almost entirely from gold variability.
In order to reduce operating costs, the bench height at Phoenix
was recently increased from six to 12 m. This change resulted in Quantification of dilution at different
better rock breakage and an increase in mill throughput but it bench heights
was not clear whether the reduced operating costs were sufficient
to offset the lost revenue due to dilution. Because of the The simulated nodes were re-blocked into Selective Mining Unit
difficulty in determining the grade of a given ore control block, (SMU) size blocks for both six and 12 m bench heights. An SMU
the conditional simulation was used to determine the relationship of 9.75 m × 9.75 m was selected for the 6 m bench height and
between bench height and misclassification. Not only can 19.5 m × 19.5 m for the 12 m bench height. A summary of the
misclassification be affected by bench height but it is sensitive to resulting tonnes, grade and net revenue is shown in Table 1. At a
cut-off grade and ore control parameters such as the number of 0.51 g/t cut-off, which is the approximate mill cut-off with no
samples used in kriging. Copper mineralisation is fairly uniform copper credit, the revenue for the 6 m bench is 62 per cent higher
in the F1 area and can be characterised as low-grade and and appears to be a substantially better choice. This significant
diffusive. While important in the analysis as it influences the difference in revenue is entirely based on geometric dilution but

TABLE 1
Comparison of 6 m and 12 m bench heights.

Cut-off 6 m bench height 12 m bench height

g/t Tonnes Grade g/t Net revenue Tonnes Grade g/t Net revenue
0.000 30 492 612 0.336 (90 095 443) 30 492 612 0.336 (90 095 443)
0.171 22 789 890 0.435 (28 282 484) 23 996 588 0.422 (35 880 458)
0.343 13 375 257 0.559 10 864 270 15 046 755 0.507 (334 741)
0.514 5 966 714 0.727 21 931 343 5 968 716 0.645 13 537 228
0.686 2 572 596 0.905 17 253 652 1 466 549 0.823 7 718 050
0.857 1 155 912 1.087 11 246 768 410 995 0.984 3 290 732
1.029 522 622 1.275 6 768 782 89 363 1.179 1 009 303

200 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 IMPACT OF BENCH HEIGHT, CUT-OFF GRADE AND PRODUCTION SPACING ON MISCLASSIFICATION IN ORE CONTROL

it does not take into account misclassification due to the
challenge of correctly predicting the block grade.
The larger the SMU size the easier it is to predict the grade of
that block, and ultimately classify it as either ore or waste, which
should increase the revenue. However, at the same time the
grades of the blocks are also decreasing due to dilution, which
lowers the revenue of the deposit. It appears that within the
normal range of cut-offs at Phoenix the dilution caused by larger
benches is much more costly than the revenue generating affects
of better ore waste classification.
Quantification of ore waste misclassification
As there is only one geologic domain (Upper Battle Fm) within
the F1 area the determination of the ore waste boundary and
subsequent classification of blocks into either ore or waste was
based entirely on the grade estimation of the SMU blocks. A
methodology was developed to determine the variability
associated with Kriged ore control estimates:
• Begin with a simulation node spacing that is one-fourth or
less of the ore control blasthole spacing (a node spacing of
1.2 m × 1.2 m was used).
• Extract simulated 4.8 m × 4.8 m blasthole patterns. In this
study 16 patterns were created using 1.2 m offsets.
• Krige the SMU size block using the extracted simulated
blastholes. This was done for the 9.75 m × 9.75 m × 6 m and
the 19.5 m × 19.5 × 12 m SMU blocks.
• Compute the tonnage and grade for each of the 16 different
models.
• Compare the models with the ‘ground truth’ simulation.
Table 2 shows the comparison of the 16 models (6 m benches)
with the simulation at the 0.5 g/t cut-off. Each of the ore control
model estimates yields a noticeably lower tonnage and grade, and
hence net revenue, than the original re-blocked simulation. This
difference between the re-blocked simulation (perfectly known
selective mining units) and the kriged ore control blocks is caused
by misclassification of ore and waste due to imperfect knowledge
from the spacing of the blasthole data, or information effect. The
poor but uniform results of the 16 models indicate that the actual
location of the blasthole data is unimportant statistically. This
implies that the continuity of the gold mineralisation is nearly
random at the current 4.8 m blasthole spacing.
One of the 16 blasthole models is plotted against the ‘ground
truth’ (re-blocked simulation) in Figure 3, showing what the
misclassification looks like. The blocks are divided into four
quadrants:
1. the lower left is waste that is modelled as waste,
2. the lower right is waste that is modelled as ore,
3. the top right is ore that is modelled as ore, and
4. the top left is ore that is modelled as waste.
This graph shows the difference between perfect selectivity,
which would be the blocks above the horizontal line
(4349 blocks at 0.72 g/t), the predicted grade, which is the
modelled blocks to the right of the vertical line (4298 blocks at
0.73) and the actual grade (4298 blocks at 0.66 g/t) which is the
true grade of the blocks to the right of the vertical line.
Appropriate number of samples for ore
control definition
The importance of misclassification in ore control is made very
clear in Table 2, which indicates that approximately half of the
possible net revenue from the F1 area is being lost. In order to
improve material classification the first step is to understand
what is causing the problem. Unfortunately, the root of the
problem is the inherent variability of the deposit, which is
believed to be too great to accurately predict block grades using
4.9 m blasthole data. With the current blasthole spacing there are
four blastholes per SMU block (6 m benches) and the optimal
number was investigated. Figure 4 shows the probability of

TABLE 2
Blasthole kriging at 4.9 m spacing versus re-blocked simulation.
Model Tonnes Grade Au (kg) Revenue Process Net revenvue
Sim 6 280 668 0.7851 4941 83 401 500 54 001 342 29 400 158
Model Tonnes Grade Au (kg) Revenue Process Net revenvue
1 5 990 306 0.6612 3965 66 918 075 51 504 804 15 413 271
2 5 843 979 0.6646 3878 65 453 325 50 246 680 15 206 645
3 5 829 681 0.6646 3874 65 395 050 50 123 744 15 271 306
4 5 684 174 0.6646 3786 63 903 000 48 872 678 15 030 322
5 5 823 085 0.6612 3857 65 100 000 50 067 038 15 032 962
6 5 813 983 0.6680 3883 65 535 750 49 988 773 15 546 977
7 5 726 926 0.6680 3832 64 684 725 49 240 253 15 444 472
8 5 765 210 0.6680 3845 64 904 175 49 569 421 15 334 754
9 5 952 819 0.6612 3935 66 417 750 51 182 492 15 235 258
10 5 760 609 0.6680 3849 64 971 900 49 529 867 15 442 033
11 6 044 293 0.6578 3986 67 278 225 51 968 990 15 309 235
12 5 910 431 0.6612 3907 65 940 000 50 818 037 15 121 963
13 5 690 320 0.6680 3803 64 197 000 48 925 516 15 271 484
14 5 871 590 0.6646 3903 65 886 450 50 484 080 15 402 370
15 5 876 538 0.6612 3894 65 726 325 50 526 622 15 199 703
16 5 730 706 0.6646 3821 64 486 800 49 272 756 15 214 044
Mean = 5 832 166 0.6642 3876 65 424 909 50 145 109 15 279 800
Stdev = 105 756 0.00328 55

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 201
W HARDTKE et al
FIG 3 - Misclassification of blocks at a 0.51 cut-off and 6 m bench height.
FIG 4 - Probability of misclassifying blocks with different number of
data points.
classifying three randomly selected blocks (6 m) correctly at a
0.51 g/t cut-off with a varying number of data points (simulated
nodes). Each of these blocks was sampled 1000 times using one
to 50 internal data points to determine the block grade. While
this figure shows blocks from the 6 m model it suggests that
when comparing the 6 m bench height which has four blastholes
per SMU block at the current blasthole spacing with the 12 m
bench which has 16 blastholes, the 12 m bench should do a much
better job of classifying the block correctly, especially routing
the ore blocks as ore instead of waste. Figure 4 implies that the
degree of misclassification can be somewhat mitigated by the use
of a sufficient number of samples. At about 16 samples the
predictive abilities of the data stabilise for the high grade
(0.80 g/t) blocks with the rate of correct classification being near
100 per cent. Low-grade (0.31 g/t) blocks are incorrectly
classified about 23 per cent of the time with 16 samples and
drops to about ten per cent by increasing the number of samples
to 50. Blocks closer to the cut-off grade see a greater benefit
202 Spectrum Series Volume 17
from an increased number of samples. These blocks are correctly
classified about 55 per cent of the time using 16 samples which
increases to close to 80 per cent when using 50 samples.
In ore control, not having a sufficient number of blasthole
assays within a SMU block to reliably predict its grade near the
cut-off is augmented by adding up to 40 of the surrounding
blastholes. This is not as desirable as having more internal data
points but still increases the probability of classifying blocks
correctly. It also tends to decrease the mean grade of those blocks
which subsequently leads to lower predicted polygon grades as
increasing the number of blasthole assays smooths the estimate.
Comparing the effects of misclassification in 6 m
and 12 m SMU blocks
In order to decrease mining costs while increasing ore production
and mill thoughput, the bench height at Phoenix was increased
6 m to 12 m. As expected, the head grade dropped due to the
added dilution of the larger SMU size. Because the lower head
grade was incompatible with the yearly production goals the
cut-off grade was increased to offset the affects of the added
dilution. However, the higher cut-off grade did not result in a
higher head grade. There can be a significant difference between
the head grade assuming perfect selectivity and the actual grade
which is a combination of ore and waste that has been
misclassified as ore. Figure 5 shows how the amount of
misclassification changes with the cut-off grade and how different
that amount is for 6 m and 12 m benches.
The difference between the 6 m actual and the 12 m actual is due
to dilution and there are very few 12 m blocks with grades
greater than about 0.6 g/t. Figure 5 also shows that the 12 m
benches significantly over predict the grade relative to the 6 m
benches. As the cut-off grade increases, the over prediction
continues to increase while the actual grade remains more or less
constant due to dilution and misclassification in the 12 m bench
case. This figure shows clearly why the increase in cut-off grade
did not result in a significantly higher grade to the mill.
The predicted grade can be calibrated so it is the same as the
actual grade with the implementation of grade capping. Figure 6
indicates that the predicted grade is the same as the actual grade
Advances in Orebody Modelling and Strategic Mine Planning I
 IMPACT OF BENCH HEIGHT, CUT-OFF GRADE AND PRODUCTION SPACING ON MISCLASSIFICATION IN ORE CONTROL

with a cap grade of approximately 3.5 g/t. This figure also shows
that the net revenue increases as the cap grade becomes higher
until it reaches about 2.5 g/t. The use of a cap grade (down to
2.5 g/t) decreases the affect of the higher grade blastholes
(smearing) and reduces the size of the ore control polygons,
resulting in an increase in revenue.

CONCLUSIONS
Misclassification of ore and waste at Phoenix can lead to a lower
than anticipated revenue regardless of the SMU size. This
misclassification can be mitigated, but not eliminated, by
increasing the number of samples used for ore control
delineation. In addition, a capping strategy should be
implemented to adjust the predicted grade to the actual grade and
to increase the net revenue.
At a cut-off grade of 0.051 g/t, the decrease in revenue because
of dilution greatly out weighs the benefit of better ore
classification when changing from 6 m to 12 m benches. While
both the 6 m and 12 m bench heights show diminishing returns
FIG 5 - Comparison of predicted, perfect and actual grades.
from an increase in cut-off grade because of misclassification,
the 12 m benches are less favourable because of extreme grade
dilution. Increases in cut-off grade will not change the mill grade
significantly.
Conversely, misclassification can be reduced by decreasing the
cut-off grade, especially when mining on the smaller bench
height. Operations should find a way to increase the
fragmentation and mill throughput for the 6 m bench height
because the cost of dilution is high. Not only is the revenue much
lower at 12 m, but 45 per cent of the potential ore is being put in
the dump at a 0.51 g/t cut-off.

REFERENCES
Deutsch, C V and Journel, A G, 1998. GSLIB, Geostatistical Software
Library and User’s Guide, second edition, 369 p (Oxford University
Press: New York).
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation,
483 p (Oxford University Press: New York).
Isaaks & Co, 2001. SAGE2001, General Purpose Software for
Variography (Users Manual), 67 p.
Phoenix Mine Geology and Mine Engineering, 2007. Mineral resource
FIG 6 - Using grade capping to adjust the predicted grade.
and ore reserve report as of December 31, 2006, Newmont Mining
Corporation internal report.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 203

A Geology-Based Joint Conditional Simulation of a Lateritic
Nickel Deposit – Quantification of Risk in Grade Tonnage Curves
and Resource Categorisation
J A Lopes1, C F Rosas2, J B Fernandes3 and G A Vanzela4
ABSTRACT
The risk quantification in grade-tonnage curves is decisive for capital
investment in mining projects. Geostatistical simulations for orebodies
can be used to obtain grade-tonnage curves, find uncertainty and risk
assessments. Applying these in multi-element deposits can be a difficult
practice, as the related attributes using the traditional co-simulation
approaches require intensive computational work and may be impractical
for use in the mineral industry. This paper presents the risk assessment
for integrating grade-tonnage curves and resources categorisation for
lateritic nickel deposits in the central region of Brazil, by joint simulation
of multiple correlated variables of interest – Ni, MgO and SiO2. The
joint simulation of these variables is based on Minimum/Maximum
Autocorrelation Factors (MAF). MAF is an approach based on spatial
decorrelates of the original variables involved, which are transformed to
non-correlated factors. The independent factors are individually simulated
and back transformed to simulate correlated deposit attributes that
reproduce the cross-correlation of the original variables. Based on this
approach, the resources are categorised honoring Ni joint simulated
results by applying the 15 per cent rule, where it is considered there will
be a statistical error <15 per cent, with 90 per cent of a confidence
interval per production period.
INTRODUCTION
Geostatistical simulation methods are used to quantify geological
risk, uncertainty in mineral deposits and assess risk in different
aspects of mining projects. Simulation methods for individual
variables are frequently efficient; however, in many applications,
reproduction of the spatial dependence and joint simulation for
multiple-variables is critical and rarely implemented (Deutsch
and Journel, 1998), for example, iron, lateritic nickel and
polimetalic deposits. Existing joint simulation methods for
multivariable deposits could be applied, like conventional
co-simulation (eg Verly, 1993; Chiles and Delfiner, 1999) and
collocated co-simulation with a Markov-type coregionalisation
(Almeida and Journel, 1994), but they are too computationally
intensive to be practical, principally when there are more then
two variables being considered. The computational inefficiency
substantially increases with the number of variables being
co-simulated. A computationally efficient alternative was
proposed by David (1988) who introduced joint simulation based
on the decorrelation of variables using Principal Component
1. MAusIMM, Senior Geologist, Votorantim Metais Ltda, Dept°
Geologia/Exploração Mineral, Edifício Esplanada, Praça Ramos de
Azevedo, 206-15° andar – Conjuntos 1530/1540, Centro, São Paulo
SP 01037-010, Brazil. Email: jose.lopes.jl1@vmetais.com.br
2. Senior Geologist, Votorantim Metais Ltda, Dept° Geologia/
Exploração Mineral, Edifício Esplanada, Praça Ramos de Azevedo,
206-15° andar – Conjuntos 1530/1540, Centro, São Paulo SP
01037-010, Brazil. Email: claudio.rosas@vmetais.com.br
3. Geologist, Votorantim Metais Ltda, Dept° Geologia/Exploração
Mineral, Edifício Esplanada, Praça Ramos de Azevedo, 206-15°
andar – Conjuntos 1530/1540, Centro, São Paulo SP 01037-010,
Brazil. Email: jorge.fernandes@vmetais.com.br
4. Senior Geologist, Votorantim Metais Ltda, Dept° Geologia/
Exploração Mineral, Edifício Esplanada, Praça Ramos de Azevedo,
206-15° andar – Conjuntos 1530/1540, Centro, São Paulo SP
01037-010, Brazil. Email: guilherme.vanzela@vmetais.com.br
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
Analysis (PCA). The effectiveness of this approach is limited,
because the PCA ignores cross-correlations at distances different
than zero. Desbarats and Dimitrakopoulos (2000) presented an
extension of the PCA approach to joint simulation of multiple
variables, replacing PCA with the Maximum/Minimum
Autocorrelation Factor (MAF). It is a factorisation method
originally developed for remote-sensing applications (Switzer
and Green, 1984). The advantage of MAF is that it is possible to
produce the uncorrelated factors following the linear model
of coregionalisation with two structures (Boucher and
Dimitrakopoulos, 2007, 2009). The MAF is an approach that
spatially decorrelates the variables involved to non-correlated
factors, which are independently simulated and back transformed
to conditional simulations that reproduce the cross-correlations
of the original variables. The joint simulation based on MAF
is shown to be an effective and relativity efficient method
of risk assessment when exiting two or more variables
(eg Dimitrakopoulos and Fonseca, 2003; Boucher and
Dimitrakopoulos, 2007).
Jointly simulated attributes of orebodies can assist resource
categorisation, a theme widely discussed in mineral resource
evaluation. Categorising resources with conditional simulation,
while considering mining production periods, is a well-known
approach in resources evaluation. This approach consists of
classifying mineral resource while considering error and
confidence intervals using the 15 per cent rule. This implies that
estimated grade, tonnage and metal content in a production
period will be within 15 per cent accuracy at a 90 per cent
confidence level. This paper presents the risk assessment of
integrating grade-tonnage curves and resource categorisation
studies for Montes Claros de Goiás (MCG) lateritic nickel
deposits by joint simulation of multiple correlated variables of
interest – Ni, MgO and SiO2, based on Minimum/Maximum
Autocorrelation Factors (MAF). The mineral resource
categorisation is applied on a Ni joint simulated result
considering error and confidence intervals per production period.
The MCG lateritic nickel deposits are located in the west portion
of Goiás state in Brazil (Figure 1). Given that the MCG has two
different types of ore, limonitic ore and saprolitic ore, this paper
was developed only considering the saprolitic ore.
MONTES CLAROS DE GOIÁS DEPOSIT AND
DATA AVAILABLE
The Montes Claros de Goiás (MCG) lateritic nickel deposits are
formed by a lateralisation process on ultramafic igneous rocks
(Dunite and Piroxenite) and are defined by chemical criteria with
different litotypes: – Soil – SO, Limonite – OX, Sarplite – SA,
Bedrock – BD and Chalcedony – CD, shown in lateritic profile in
Figure 2. The saprolite and limonite are mineralised types where
saprolitic ore is destined to the Ferronickel (FeNi) process and
limonitic ore is destined to High Pressure Acid Leaching (HPAL)
process. There are 26 768 one metre composites samples from
reverse circulation drilling (RC), where 20 474 samples are
limonitic ore and 6294 samples are saprolitic ore. Table 1 and
Figure 3 show the statistics resume and histogram for Ni, MgO
and SiO2.
Spectrum Series Volume 17 205
J A LOPES et al

FIG 1 - MCG deposit location.

FIG 2 - MCG lateritic profile.

206 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 A GEOLOGY-BASED JOINT CONDITIONAL SIMULATION OF A LATERITIC NICKEL DEPOSIT

FIG 3 - Histogram from Ni, MgO and SiO2.

Λ1 = Q1B Q1T (3)

TABLE 1
Statistics of one metre composites.
and Q2 is the matrix of eigenvectors from decomposition:
Ni (%) MgO (%) SiO2 (%)
Number sample 6294 6294 6294 Q2M(Δ)Q2T = Q2{½[[Γ y(Δ)]T + [(Γ y(Δ)]]}Q2T (4)
Mean 1.08 18.31 35.85
Where matrix Γ y(Δ) is an symmetric variogram matrix at lag
Median 0.96 17.95 35.15 distance Δ for regular PCA factors:
Std deviation 0.58 6.18 7.16
Variance 0.34 38.14 51.24 Y(x) = Z(x) A
Kurtosis 0.68 -0.69 2.99
where:
Asymmetry 0.89 0.16 1.16
Maximum 4.27 38.70 69.2 A = Δ-1/2
Minimum 0.06 1.28 10.8
Several Δ lag distances may be used for values lower than the
range, and the resulting eigenvectors averaged.
METHODOLOGY The joint simulation of multiple correlated variables using the
MAF approach process steps are below:
Joint simulation using maximum/minimum • normalise the available variables,
autocorrelation factors (MAF) • use MAF transformation to generate non-correlated factors
(MAF),
Multi-elements are represented by multivariate stationary and
ergodic vector random function. Considering a multivariate l • produce the variogram for each factor,
dimensional, Gaussian, stationary and ergodic spatial vector • conditionally simulate each factor,
random function:
• validate the simulation of factors,
Z(x) = [Z1(x),…, Zl (x)] T • back transform the simulated factors to normal score variables,
• de-normalise the simulated variables, and
The Maximum/Minimum Autocorrelation Factors are defined
as the l of the orthogonal linear combinations:
• validate final results.

Y(x) = aiTZ(x), i = 1,…, l Resource categorisation using the 15 per cent

of the original multivariate vector Z(x). MAF are derived
assuming Z(x) is represented by a two-structure linear model of
coregionalisation. The MAF transformation can be written as:
Y(x) = AMAFZ(x)
and the MAF factors are derived from:
AMAF = Q2Λ1-1Q1
Where the eigenvectors Q1 and eigenvalues Λ1 are obtained
from decomposition of the multivariate covariance matrix B of
Z(x) at lag distance:
rule
The mineral resource categorisation using the 15 per cent rule is
estimated within 15 per cent of the true value, with 90 per cent
confidence in the production period. This implies that the
estimated grade, tonnage and metal content in a production
(1) period will be within 15 per cent accuracy at a 90 per cent
confidence level, ie nine out of ten times production periods will
have an estimation error less than 15 per cent. According to
Dohm (2004), who describes an application of this resource
categorisation criterion using threshold values, the categorisation
(2) criteria are as follows:
1. calculate the error with 90 per cent confidence:
Error = 1.645 CoV / n

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 207
J A LOPES et al

where: TABLE 2
Transformation matrix.
CoV = (Std Deviation)/Mean
-0.546 0.679 0.471
n = number of blocks mined per period 0.083 0.424 -0.088
2. Change support, ie block to production quarter and block to 0.833 0.598 0.077
one year production period – calculate the variability
reduction factor:
3. FPQ = CoVBlock / CoVPQ
FPY = CoVBlock / CoVPY
where:

CoVBlock = CoV for blocks of the same size used for estimation

CoVPQ = coefficient of variation of quarter production period

CoVPY = coefficient of variation of year production period

4. Calculate the measured and indicated threshold for

classification;

Measured (Threshold) = (FPQ × 0.15)/1.645

Indicated (Threshold) = (FPY × 0.15)/1.645

5. Classifying resources:

Measured = CoVBlock ≤ Measured (Threshold)

Indicated = Measured (Threshold) < CoVBlock ≤

Indicated (Threshold)

Inferred = Indicated (Threshold) > CoVBlock

After resource classification, it is an important practice to

validate the classification with geology and draw the final
measured, indicted and inferred limits.

JOINT SIMULATION OF Ni, MgO AND SiO2

Normal-score transformation
The normal-score transformations are based on rank ordering of
data and decrease the influence of outliers. Transform the Ni,
MgO and SiO2 original data distribution to normalised data
distribution. This assists in the calculation of the variogram and
covariance matrix.
MAF transformation and variogram analysis
The MAF transformation is calculated by multiplying a vector of
variables Ni, MgO and SiO2 by a vector loading from the
transformation matrix AMAF (Equations 1, 2, 3), and used to
generate the three min/max autocorrelation factors for each
variable (Table 2). The lag distance Δ used in transformation
Equation 4 is 15 m taken from variogram analysis on MAF
values, which demonstrates decorrelation in cross variogram
between calculated MAF values (Figure 4). The variograms of
saprolitic ore are performed and fitted for each MAF by
traditional spherical variogram (Figure 5). The MAF variograms
are subsequently used for the simulation.
FIG 4 - Ni, MgO and SiO2 MAF cross-variogram.
simulation (Dimitrakopoulos and Luo, 2004). The simulation is
performed on a grid of 117 620 nodes within geological limits.
Twenty realisations and validations are run for each MAF
reproducing histogram and variogram. The validation of the
MAF simulations is not presented here. Only the final validation
in the data space is presented.
MAF back transformations
The realisations of each MAF were back transformed to
simulated normal score variables, using the corresponding
inverse matrix of the MAF loadings in Table 2. The normal score
for each variable, Ni, MgO and SiO2, were back transformed to
data space. Three realisations and the data of Ni grade
distribution are shown in Figure 6.

Conditional simulation of MAF Validation of joint Ni-MgO-SiO2 simulations results

The conditional simulation is performed independently for each Several validation checks are performed on the results of joint
MAF, using a sequential algorithm based on sequential Gaussian Ni, MgO and SiO2 simulations using the MAF transformation.

208 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 A GEOLOGY-BASED JOINT CONDITIONAL SIMULATION OF A LATERITIC NICKEL DEPOSIT
FIG 5 - Ni, MgO and SiO2 MAF variogram.
The validation involves histograms, variograms and cross-
variograms comparison between simulation and data. Figure 7
shows histograms comparison data versus simulation (one
realisation randomly chosen). Figure 8 shows variogram and
cross variogram comparison of data versus simulation. All
validation results suggest that the reproduction of the original
data by the joint simulations is good.
RISK ASSESSMENT AND RESOURCES
CATEGORISATION
Lateritic nickel deposits have limitation processes caused by
SiO2, MgO, Fe and Al2O3 grades, depending on what process
type is to be used. Nickel production by a Ferronickel (FeNi)
process is strongly influenced by the SiO2/MgO ratio. The
SiO2/MgO must be up to 1.7 without copper coolers conditions
and up to 2.0 with copper coolers conditions. A SiO2/MgO ratio
more then 2.0 can generate operational furnace problems, and
depending on the persistence of higher SiO2/MgO ratios, the
project can became unviable. The grade-tonnage curves
Advances in Orebody Modelling and Strategic Mine Planning I
simulation results of Ni, MgO and SiO2 in MCG saprolitic ore
(Figure 9), show 20 possibilities of tonnage, Ni mean grade and
SiO2/MgO ratio considering Ni cut off. Based on these results, it
is possible quantify the risk of the Ni mean grade to be too low
and/or the SiO2/MgO ratio to be too high, supporting important
decisions to the project.
Based on the joint simulation results of Ni, MgO and SiO2 in
the MCG saprolitic ore, the resources were categorised
considering statistical error and confidence intervals, honoring
each step described earlier and using the 15 per cent rule.
Resources categorisation using the 15 per cent rule criteria implies
that estimated grade, tonnage and metal content in a production
period will be within 15 per cent accuracy at a 90 per cent
confidence level. The MCG saprolitic ore has blocks that are 25 ×
25 × 1 m3. These blocks were changed to support a quarter period
production block size of 200 × 200 × 7 m3 and to a one year
production period block size of 200 × 200 × 28 m3 (Figure 10).
Figure 11 shows the distribution error at a 90 per cent confidence
for blocks, quarter and yearly production periods. The quarter year
block size was used to calculate the Measured resources and the
one year block size was used to calculate the Indicated resources.
Spectrum Series Volume 17 209
J A LOPES et al

FIG 6 - Ni grade distribution for three realisations.

FIG 7 - Validation – histogram comparison.

210 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 A GEOLOGY-BASED JOINT CONDITIONAL SIMULATION OF A LATERITIC NICKEL DEPOSIT
FIG 8 - Validation – variogram and cross-variogram comparison.
The MCG resources categorisation is then:
Measured = CoVBlock ≤ Measured (Threshold)
Indicated = Measured (Threshold) < CoVBlock
≤ Indicated (Threshold)
Inferred = Indicated (Threshold) > CoVBlock
After resources were classified, some isolated measured blocks
were considered indicated. Figure 12 shows a plan view with
Measured, Indicated and Inferred resources distribution results of
the categorisation criteria described above.
CONCLUSIONS
The risk analysis in grade-tonnage curves on multi-element
deposits shows it is applicable in lateritic nickel deposits, or for
any multi-element deposit, and should be considered in feasibility
studies and in supporting important decisions concerning the
project. The approach applied in this study presented an efficient
joint conditional simulation using minimum/maximum auto-
correlation factors, which simplifies joint simulation in
multi-element mineral deposits by decorrelating variables to
independent factors that are simulated independently. The
comparison between simulated results versus original data shows
excellent validation. The resources categorisation presented in
this study, classifying Measured, Indicated and Inferred
resources, while honoring the error and confidence interval in a
production period, is an important approach to support short and
long term production planning.
The MCG deposit has two types of orebodies: saprolitic, and
limonitic ore. This study focuses only on saprolitic ore treated
Advances in Orebody Modelling and Strategic Mine Planning I
with a Ferronickel (FeNi) process. Considering the FeNi process
requirement, the results of twenty realisations show acceptable
SiO2/MgO ratio for any nickel cut off considered.
REFERENCES
Almeida, A S and Journel, A G, 1994. Joint simulation of multiple
variables with a Markov-type coregionalisation model, Mathematical
Geology, 26(5):565-588.
Boucher, A and Dimitrakopoulos, R, 2007. A new efficient joint
simulation framework and application in a multivariable deposit, in
Orebody Modelling and Strategic Mining Planning, second edition
(ed: R Dimitrakopoulos), pp 345-354 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Boucher, A and Dimitrakopoulos, R, 2009. Block simulation of multiple
correlated variables. Mathematical Geosciences, 41(2):215-237.
Chiles, J P and Delfiner, P, 1999. Geostatistics Modelling Spatial
Uncertainty, 695 p (John Wiley and Sons: New York).
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
Reserve Estimation, 216 p (Elsevier: Amsterdam).
Desbarats, A J and Dimitrakopoulos, R, 2000. Geostatistical simulation
of regionalised pore-size distribution using min/max autocorrelation
factors, Mathematical Geology, 32(8):919-942.
Deutsch, V C and Journel, A G, 1998. Gslib: Geostatisitcal Software
Library and User's Guide, second edition, 369 p (Oxford University
Press: New York).
Dimitrakopoulos, R and Fonseca, M B, 2003. Assessing risk in
grade-tonnage curves in a complex copper deposit, northern Brazil,
based on an efficient joint simulation of multiple correlated variables,
Application of Computers and Operations Research in the Minerals
Industries, pp 373-382 (Southern African Institute of Mining and
Metallurgy: Marshalltown).
Dimitrakopoulos, R and Luo, X, 2004. Generalised sequential Gaussian
simulation on group size and screen-effect approximations for large
field simulations, Mathematical Geology, 36(5):567-591.
Spectrum Series Volume 17 211
J A LOPES et al

FIG 9 - Grade-tonnage curves.

FIG 10 - Change nodes to three month and year production supports.

212 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 A GEOLOGY-BASED JOINT CONDITIONAL SIMULATION OF A LATERITIC NICKEL DEPOSIT

FIG 11 - Error percentage with 90 per cent confidence – plan view distribution.

FIG 12 - Resources categorisation – plan view distribution.

Dohm, C, 2004. Quantifiable mineral resource classification: A logical Verly, G W, 1993. Sequencial Gaussian Co-simulation: A Simulation
approach, in Geostatistics Banff – Book Series Quantitative Geology Method Integrating Several Types of Information in Geostatistics (ed:
and Geostatistics, 2:333-342 (Springer: New York). A Soares), 5:543-554 (Kluwer Academic Publishers: Dordretcht).
Switzer, P and Green, A A, 1984. Min/max autocorrelation factors for
multivariate special imagery, technical report 6, Stanford University,
Department of Statistics, 10 p.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 213

Resource/Reserve Classification with Integrated Geometric and
Local Grade Variability Measures
R Dimitrakopoulos1, M Godoy2 and C L Chou3
ABSTRACT
Categorisation of mineral resources and ore reserves as Measured,
Indicated, Inferred and Proven or Probable respectively, is a requirement
for both public reporting as well as internal company assessments of
mining assets. From the understanding of mineral resources to estimation
of ore reserves, geological uncertainty impacts heavily upon resource/
reserve classification. Resource/reserve classification is traditionally based
on geometric criteria, as by the number of samples or drill holes within a
volume around the deposit blocks being classified, or alternatively, some
form of a geostatistically-derived estimated variance. These criteria are
relatively easy to understand and implement – however, they do not
account for the local variability of grades within a deposit, leading to
misrepresentation of the confidence of estimated grades within given
volumes.
This paper proposes an approach founded on a stochastic simulation
framework for quantifying geological uncertainty and risk in five steps:
generation of suitable realisations of the deposit at the appropriate block-
support scale, derivation of conditional cumulative distribution functions
(ccdf) of block grades or attribute selected, utilisation of the conditional
coefficient of variation (CCV) of the grade ccdf within a block as the
uncertainty measure for resource classification of each block in the
orebody model, classification of resources/reserves, and evaluation and
production of risk profiles for estimates. This approach integrates both the
traditionally-used geometric considerations, in the sense of data locations
and the in situ variability of grades. The applications in an alluvial diamond
deposit, epithermal gold lode and a porphyry copper deposit show that
there are natural groupings of the CCV of mining blocks that allow rational
categorisation of resources or reserves. The proposed approach is not only
transparent but also provides a technically defendable reporting of
resource/reserve class- ification in a geostatistical context and may support
better investment decision making.
INTRODUCTION
Classification of mineral resources as Measured, Indicated and
Inferred and ore reserves as Proven or Probable is required for
public disclosure as detailed in various codes such as JORC, CIM,
PERC, or SAMREC. In addition, categorisation of resources and
reserves is an important contributor to a company’s assessment of
mining assets and their grade-tonnage curves. It is widely accepted
that geological uncertainty and risk heavily impact upon the
classification of resources and reserves, starting from the
understanding of mineral resources and exploration programs to
the ‘ore reserve’. Reliability, confidence, transparency and
technical ‘defendability’ of classification are critical. Considering
the worth of mining companies, benefits to shareholders, reporting
to financial institutions, and attracting the investment community
to decision-making in the order of billions of dollars, there is a
genuine need to understand and quantify geological uncertainty
1. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montreal QC H3A 2A7, Canada.
Email: roussos.dimitrakopoulos@mcgill.ca
2. MAusIMM, Golder Associates Santiago, Av 11 de Septiembre 2353,
Piso 2, Proviendencia Santiago, Chile. Email: mgodoy@golder.cl
3. COSMO – Stochastic Mine Planning Laboratory, Department of
Mining and Materials Engineering, McGill University, FDA
Building, 3450 University Street, Montreal QC H3A 2A7, Canada.
Email: lulu.chou@mail.mcgill.ca
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
and risk underpinning resources/reserves. State-of-the-art
approaches that quantify geological uncertainty and risk can be
used to standardise resource/reserve classification, support better
investment decision making, enable technically-defensible stock
exchange reporting and compliance with industry codes and
provide greater assurance to investors and protect the public.
Traditionally, resource classification has been based on
geometric considerations and a block-by-block classification
using criteria linked to the number of samples with-in a volume
or drill hole densities, or alternatively, a geostatistically-derived
estimation variance. These criteria are relatively simple to
implement as well as being easily understood and transparent.
However, they have a main drawback as they do not account for
the local variability of grades within a deposit – this leads to
misrepresentation of the confidence of estimated grades within
given volumes and of the resulting resource/reserve
classification.
Past efforts to improve classification techniques focused on the
development and use of alternative versions of grade estimation
(kriging) variances, where classification patterns are based on the
size of the sampling grid (eg Saburin, 1983; Froideveau, Roscoe
and Valiant, 1986; David, 1988; David and Dagbert, 1977; Arik,
2002 and others). As also noted above, these formulations may
account for possible increases in precision as the sampling grid
size decreases – however, they do not incorporate uncertainty
introduced by local variability in the actual data as eloquently
presented by Ravenscroft (1992). Figure 1 illustrates this point,
showing two identical configurations of a block and four drill
holes. On the right, the grade of the material sampled is highly
variable, whereas on the left block, the grades are much less
variable. If this were the same zone with the same variogram, the
estimation variances would be the same, thus providing a false
level of uncertainty assessment and classification of the blocks.
Estimation variances do not adequately address the question of the
precision of global resource estimates and even more so for global
resources above a cut-off (David, 1988), as would be the case if
any other geometric measure was used. Simulation-based
approaches have been suggested given the issues discussed above
(Audet and Ross, 2007; Wawruch and Betzhold, 2005; Dohm,
2005; Duggan and Dimitrakopoulos, 2005; Murphy et al, 2004)
and need further consideration.
The shortcomings of the conventional reserve/resource
classification may be explored in an example from a gold mine.
Reserves classified as Measured, Indicated and Inferred are based
primarily on the number of samples within a distance from a block
being geostatistically estimated and consideration of the spatial
grade continuity within identified gold lodes. Figure 2 shows part
of a lode where blocks have been classified to reserve categories
according to the mine’s established approach. Although this
FIG 1 - Estimation for identical sampling configurations with
different local grade variability and uncertainty.
Spectrum Series Volume 17 215
R DIMITRAKOPOULOS, M GODOY and C L CHOU
FIG 2 - Traditional reserve classification in a lode of a gold mine
(left) and probabilities for the blocks to be ore (right).
approach is reasonable, relatively simple, transparent, compliant
with regulations and technically defensible, the traditional
approach shows shortcomings when the classification is compared
to probabilities of blocks being ore (see Figure 2). Large parts of
the lode that are classified as Indicated have over 80 per cent and
90 per cent probabilities of being ore – thus it seems inconsistent
not to report them as Measured. Similarly, blocks which have been
classified as waste have, in fact, over 50 per cent probability of
being ore. Although the analysis above is confined to limited
aspects (for example and as per the JORC, one must also
accurately estimate the grade), it illustrates how conventional
classification criteria do not include direct measures of geological
(grade) uncertainty apart from those related to the spatial
configuration of the drill holes. As a result, the conventional
approaches may lead to different classification of Mineral
Resources compared to using classifications which integrate the
actual in situ spatial variability of the orebody.
In the following sections, stochastic simulation techniques are
shown to provide critical information for resource/reserve
classification complementing resource estimates, with the
understanding of the multifaceted nature of the classification
process and the role of the competent person. The techniques
model uncertainty in estimated grades of orebody models and lead
to an assessment of geological uncertainty as a function of both in
situ variability and drilling density. In addition, the simulation
framework allows for the integration of simulated geological
alternatives with simulated grades and possible errors in assay
data, drill hole locations and so on. The next section begins by
briefly discussing limits of traditional classification approaches.
Subsequently, an alternative methodology is outlined, based on
simulated conditional distributions of block- support grades and
measures of their uncertainty for an orebody model. Measures of
uncertainty are detailed next using an example from an alluvial
diamond deposit. Resource/reserve classification in the same
deposit, an epithermal gold lode and a porphyry copper deposit
elucidate the related aspects and characteristics of the approach.
Comments and conclusions follow.
RESOURCE/RESERVE CLASSIFICATION AND
MEASURES OF LOCAL UNCERTAINTY
Conditional simulation enables a resource/reserve class- ification
methodology that simultaneously considers data density and
orebody continuity as well as in situ variability. In this section,
key aspects of the methodology are described, including
appropriate measures of uncertainty to be used for resource/
reserve classification. The methodology is illustrated using a
diamond-bearing deposit in south-western Africa.
216 Spectrum Series Volume 17
Classification methodology
A resource classification methodology is outlined in Figure 3 and
consists of five steps:
1. generation of suitable realisations of the deposit at the
block-support scale considered appropriate,
2. derivation of conditional cumulative distribution functions
(ccdf) of block grades or attribute selected,
3. calculation of uncertainty measure(s) to be used for
resource classification based on the ccdf of each block in
the orebody model,
4. classification of resources/reserves, and
5. evaluation and production of risk profiles for estimates.
Measures of local spatial uncertainty
A key part of the classification methodology is the use of
summary statistics and measures of uncertainty which are derived
from the conditional cumulative distribution function of the grade
Z(u) of a block at location u within the deposit (Figure 4). Basic
measures include the conditional variance, conditional coefficient
of variation, inter percentile ranges and probability interval.
The conditional variance (CV) measures the spread of the ccdf
around its mean value and is given by:
K +1
CV ( u) = ∑[z k − zE* ( u)]2 [ F ( u; zk ) − [ F ( u; zk − 1 ) ] (1)
k= 1
where:
zk k=1,…K are K threshold values discretising the range of
variation of z values
zk is the mean of the class zk-1,(zk-1,zk] which in case
of a within-class linear interpolation model
corresponds to zk = (zk-1 + zk)/2
K +1
zE* ( u) = ∑ z [ F ( u; z ) − F ( u; z
k k k − 1 )] is the discrete sum
k= 1 approximation to the expected value of the ccdf
The conditional coefficient of variation (CCV) corresponds to
the conditional standard deviation divided by the mean or relative
standard deviation. It expresses variability as a percentage of the
mean, and is calculated as:
K +1
∑[z k − zE* ( u)]2 [ F ( u; zk ) − F ( u; zk − 1 )]
k= 1
CCV ( u) = (2)
zE* ( u)
The CCV better discriminates zones in terms of sample density
(or availability of conditioning information). An advantage of the
CCV over the CV is that it is expressed directly as a percentage of
the mean and therefore filters out proportional effects.
The inter quartile range (IQR) is defined as the difference
between the upper (q0.75) and the lower (q0.25) quartiles of the
distribution:
IQR( u) = q0 .75 ( u) − q0 .25 ( u) = F −1 ( u; 0.75) − F −1 ( u; 0.25) (3)
Similarly, the generalisation to any inter percentile range
(IPR), such as a 90 per cent confidence interval, is defined as:
IPR( u) = q0 .95 ( u) − q0 .05 ( u) = F −1 ( u; 0.95) − F −1 ( u; 0.05) (4)
Unlike the CV and CCV, IPR and its variations are not affected
by the magnitude of grades in the ccdf and are thus not sensitive
Advances in Orebody Modelling and Strategic Mine Planning I
 RESOURCE/RESERVE CLASSIFICATION WITH INTEGRATED GEOMETRIC AND LOCAL GRADE VARIABILITY MEASURES
FIG 3 - Diagrammatic representation of a simulation-based resource classification methodology.
FIG 4 - A conditional cumulative distribution function (ccdf)
showing summary statistics and uncertainty measures for the
thickness of a 100 × 100 m2 block in a diamond deposit. The ccdf
is generated by calculating the associated probabilities for a given
block from multiple simulated realisations.
to outliers. The IPR may also be scaled by the mean or median of
the corresponding ccdf. Related measures of uncertainty are the
probability that a block grade, Z(u) is above, below or within a
probability interval. Figure 4 shows related summary statistics and
uncertainty measures for the ccdf of thickness of a 100 × 100 m2
block. The ccdf itself provides the model of uncertainty for the
value of the thickness attribute for a given block at location u
within the deposit. Each discrete point in the ccdf corresponds to
a simulated value. The ccdf is a cumulative histogram of all
simulated values assigned to the simulation realisations of the
deposit. A continuous function is interpolated between the
discrete points to enable the assessment of probabilities for any
ccdf value.
Figure 5 shows the CV, conditional mean, CCV and IQR for the
same size of thickness blocks over a part of a diamond deposit.
The values are generated from 50 simulation realisations and drill
Advances in Orebody Modelling and Strategic Mine Planning I
hole locations are shown on the figure by triangles. The CV shows
good correlation between regions of dense sampling and low
conditional variance. However, this is not necessarily true for areas
of high thickness. The presence of a proportional effect is clearly
seen when comparing the CV and mean block thickness values,
suggesting that that the CV may not be a suitable measure of local
uncertainty. Similarly to the CV, the IQR shows good correlation
between regions of dense sampling and lower IQR values as well
as a higher uncertainty about the higher thickness values.
Generally, for highly skewed distributions the IQR may appear
more attractive than the CV as the IQR is a relative measure that
does not use the mean. A major drawback of the IQR is that it
ignores the internal distribution of probability densities, leading to
the over-representation of uncertainty. The CCV discriminates
better than the CV and IQR in terms of sample density (available
conditioning information) and, more importantly, does not exhibit
a proportional effect. As a result, it is more suitable for resource
classification.
Resource/reserve classification measures
The uncertainty measures in the previous section can be used to
define resource/reserve classification criteria. These criteria
require the use of threshold values, similarly to past efforts, where
thresholds were applied to estimation variances (eg review in
Sinclair and Blackwell, 2000). Threshold values may reflect the
error tolerance that is acceptable locally for the block estimates as
well as globally for the resource estimate. The block estimate
herein corresponds to the mean value (or e-type estimate). It is
possible to categorise resources with block precision and using
resource models derived from multiple simulation realisations to
obtain the uncertainty in global resources or resources within
production increments of interest in each category.
A first confidence interval based on the CCV involves fixed
threshold CCVs that are compared to the CCV. This is equivalent
to comparing the confidence interval about the mean of ± the
standard deviation against a threshold confidence interval
defined as a proportion of the mean. The CCV classification
criterion are given by:
Spectrum Series Volume 17 217
R DIMITRAKOPOULOS, M GODOY and C L CHOU

FIG 5 - Plan view from left to right, the conditional variance (CV); conditional mean (CM); conditional coefficient of variation (CCV); and inter
quartile range (IQR) for 100 × 100 m2 block thickness in a diamond deposit. Note that denser drilling does not ensure less uncertainty.

⎧ Measured if CCV < Threshold 2 CASE STUDIES

⎪
CCVClassification = ⎨ Threshold 2 ≤ Indicated if CCV < Threshold 1(5)
⎪ Inferred if CCV ≥ Threshold 1 A simple example in a diamond deposit
⎩

As an example, consider a CCV classification with a threshold
CCV of 0.5. In this case, all blocks for which the confidence
interval bounded by ± the standard deviation is over 50 per cent
of the value of the mean will be classified as inferred. The
remaining blocks are classified as Indicated or Measured. The
selection of threshold values is deferred until a later section.
A second confidence interval based on 90 per cent probability
(a relative percentile range or RPR) is a confidence interval about
the mean that is compared with a threshold confidence interval
defined as a proportion of this mean. The CCV classification is
defined based on the standard deviation, here the confidence
interval is defined considering the relative position of the mean
to the fifth and 95th percentiles, termed negative and positive
differences. The symmetric confidence interval used in this
criterion is:
⎡ z*E ( u) − F −1 ( u;0. 05) z*E ( u) − F −1 ( u;0. 95) + ⎤
Confidence interval = ± ⎢ *
w− +
⎢ zE ( u) z*E ( u)
⎣ ⎦
where:
w- and w+ are weights for the relative differences to the fifth and
95th percentiles, respectively. The weights facilitate asymmetric
confidence intervals and are determined based on the probability
intervals:
[ 0.05 − F ( u; zE* ( u))] [ 0.95 − F ( u; zE* ( u))]
w− = w+ =
0.9 0.9
Note that setting w- and w+ to 0.5 corresponds to defining the
confidence interval as ± the average between the negative and
positive differences. The classification criterion is:
⎧ Measured if Threshold < Confidence interval 2
⎪
RPR Classification = ⎨ Confidence interval 2 ≤ Indicated if threshold < Confidence interval 1
⎪ Inferred if threshold ≥ Confidence interval 1
⎩
For an alluvial diamond deposit where the important classification
parameter is thickness, Figure 6 shows a plan view of the resource
classification as Indicated or Inferred using three classification
criteria (CCV with a CV threshold of 40 per cent, RPR with a
90 per cent confidence interval and a single (for simplicity)
threshold of 40 per cent, and CCV with a CV threshold of 50 per
cent). Table 1 details the potential classified resource in this
diamond deposit and related estimates with their uncertainty.
Using the CCV classification, both the volume classified as
Indicated and the associated uncertainty, as measured by the
differences in maximum and minimum volumes, increase with
threshold, as expected. The increase in volume is considerable
while the increase in uncertainty in relation to the expected value
is relatively small. In global terms, the uncertainty in the volume
classified as Indicated is relatively low. The results indicate that in
this study the 40 per cent threshold may be considered conservative.
For inferred resources, as the threshold increases the volume
decreases and associated relative uncertainty increases. The
w ⎥(6) decrease in volume results from blocks moving to the Indicated
⎥ category. The increase in uncertainty is expected as it reflects the
higher uncertainty related to higher threshold values. The RPR
classification when compared to the CCV classification approach,
classifies considerably fewer blocks as Indicated at all thresholds
and does not represent a good assessment of variability. As with
the IQR measure discussed earlier, the RPR classification ignores
the internal distribution of probability densities and leads to the
over-representation of risk. It is therefore very conservative in
classifying blocks. The results presented in Table 1 show that the
uncertainty analysis on the resource based on three different
(7) threshold values, may provide an indication as to which threshold
value could be more appropriate for the deposit. Understanding
the meaning of the threshold value is critical in applying the
proposed classification criteria. Mining and processing parameters
may further be considered in addition to external considerations
that may be taken on board, relevant to project economics or a
company’s risk aversion level. Note that the analysis herein
corresponds to propagating local block uncertainty to the
uncertainty in the global resource. Selection criteria of threshold
values based on deposit characteristics are discussed next.

218 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 RESOURCE/RESERVE CLASSIFICATION WITH INTEGRATED GEOMETRIC AND LOCAL GRADE VARIABILITY MEASURES

TABLE 1
Resource classification of 100 100 m2 thickness blocks based on CCV and RPR classification criteria using 40 per cent, 50 per cent
and 60 per cent CV thresholds.

Threshold CCV classification RPR classification

40% 50% 60% 40% 50% 60%
Category Indic Infer Indic Infer Indic Infer Indic Infer Indic Infer Indic Infer
Average volume (M m3) 3.15 3.91 5.26 1.80 6.67 0.46 0.45 6.61 2.06 5.00 3.94 3.13
Max volume (M m3) 3.02 4.27 5.66 2.00 7.17 0.55 0.50 7.16 2.19 5.45 4.21 3.43
3)
Min volume (M m 3.38 3.60 4.96 1.64 6.25 0.36 0.40 6.25 1.91 4.70 3.74 2.89

against the CCV for 25 block support realisations of the orebody.

FIG 6 - Resource classification using the CCV classification
criterion with a CV threshold of 40 per cent (left); RPR
classification within 90 per cent confidence interval and a CV
threshold of 40 per cent (middle); and CCV classification using a
CV threshold of 50 per cent (right).
Selecting thresholds at a gold mine
The example of an open pit gold mine in Western Australia is used
to further explore CCV based classification. Issues addressed
include the selection of threshold values for reserve classification.
The example is comprised of a set of epithermal gold lodes in a
zone of the deposit which are to be classified as Measured,
Indicated or Inferred resources. Figure 7 shows tonnage of ore
above cut-off (1 g/t) and average grade above cut-off plotted
The figure shows a general and distinctly increasing spread (risk)
in ore tonnage and average grade as the CCV increases, which is
particularly strong in the ore tonnages. This trend is also seen in
Table 2 detailing CCV and summary statistics for metal content in
the gold lode. The characteristics of the spread and slope of the
curves in Figure 7 and values reported in Table 2 can be used to
guide selection of thresholds for classification. For example,
Figure 7 suggests that there is little risk for a CCV up to around
0.65 as the related tonnage above the cut-off curves are mostly
overlapping and the average grade above cut-off appears to have
stabilised. In addition, the figure shows that after a CCV of about
1.1, the spread of tonnages and grades is at its greatest and
remains stable for higher CCV values. Hence, the higher CCVs
can be selected for the classification of the high-risk resource.
In an alternative analysis, Table 2 shows a CV with a
minimum of 4.2 per cent at the 0.65 CCV threshold, suggesting
that this threshold corresponds to the most certain measure of
metal quantity and would be a reasonable threshold to
differentiate between Measured and Indicated reserves. The
differentiation between Indicated and Inferred reserves may be
based on a CV of about 6.7 per cent or CCV of 1.1, which
identifies the highest end of risk. At a higher threshold only a
small amount of additional metal would be delivered. For a CCV
value between 0.65 and 1.1, both metal quantity and uncertainty
increase substantially. Hence, there is a trade-off between gains
in metal quantity and increase in risk when differentiating
between Indicated and Inferred reserves. Note that Table 2 allows
straightforward reporting of average metal for each category as
well as a confidence interval, minimum, maximum and
proportion of range, as in Table 3 (the ± in the table are in terms
of ranges between maximum and minimum simulations). Also
note that the tolerance used in this case study could be lower
resulting in possibly expanding the Measured resources. Table 3
provides the detailed reporting of the reserve classification. An

FIG 7 - Average grade above the 1 g/t cut-off (left) and tonnage of ore above cut-off (right) plotted versus the CCV for 25 realisations
of a gold load.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 219
R DIMITRAKOPOULOS, M GODOY and C L CHOU

TABLE 2 spread (risk) in ore tonnage as the CCV increases. Figure 8

suggests the presence of little risk for a CCV up to about 0.45 as
CCV and summary statistics for metal content in a gold load.
the related tonnage above the cut-off curves are mostly
CCV Min Max Range Average CV Increase overlapping. The spread of the same curve increases up to a CCV
metal metal metal (%) in metal of 0.85, when the spread appears to reach its maximum and
(oz) (oz) (oz) quantity stabilises at that maximum spread. These two values of the CCV
0.35 576 1151 575 812 17.0 can then be used as the threshold values chosen for resource
classification.
0.40 2925 4645 1720 3849 10.3 3037
0.45 6721 9834 3113 8840 7.8 4991 Comments
0.50 14 245 20 000 5755 18 120 6.3 9280
It is important to note that the thresholds used in the above
0.55 29 146 36 105 6959 33 353 4.4 15 233 examples will change from deposit to deposit, different geological
0.60 48 050 58 724 10 674 53 916 4.3 20 563 zones and different commodities. However, the process of
generating CCV values will remain the same and the relatively
0.65 72 354 83 978 11 624 77 909 4.2 23 993
distinct brakes in the various curves and corresponding tables will
0.70 97 457 116 560 19 102 107 747 4.7 29 838 be present. In addition, it is important to reiterate that unlike
0.75 127 570 155 776 28 206 140 369 5.3 32 622 traditional approaches, classification of the blocks using any of the
proposed measures generated through use of simulations reflect
0.80 151 644 187 932 36 288 168 126 5.4 27 757
both data density and local grade variation. Figure 9 shows the
0.90 191 436 238 876 47 440 212 314 6.0 44 188 CCV classification of blocks in a section of the gold lode
1.00 212 766 270 189 57 423 238 691 6.4 26 377 discussed above, along with the available drilling data. Although
drilling density varies, the classification does not always show a
1.10 223 778 286 719 62 941 251 968 6.7 13 277 strong correlation with drilling density. For example, a densely
1.20 228 104 295 563 67 459 258 527 6.9 6559 drilled area may not always be classified as Measured, as
1.30 229 903 298 081 68 178 261 089 6.9 2562 traditional classification approaches may imply.
1.40 230 305 299 453 69 148 262 317 7.0 1227
Blocks reported in resource classification should be of the
same support size. If this is not the case, support-independent
1.50 230 549 299 779 69 230 262 641 7.0 324 measures may be derived. A known approach is to standardise
the CCV of a block of a given support size using the square root
of the corresponding dispersion variance (eg Sabourin, 1983).
Another practical comment is that the number of simulations
TABLE 3 required in this approach depends on the sensitivity of the results.
Reporting of classification and evaluation of ore reserves for If additional simulations do not effect the overall classification,
metal, tonnes and average grade. they are not required. All classification approaches may be
Category CCV Tonnage Average Metal
improved by choosing a rational spatial grouping of classified
(t) grade (g/t) (oz) blocks in a category to avoid their unrealistic dispersion. Last but
not least, the ability to efficiently simulate large deposits and
Measured <0.65 790 ± 23 2.80 ± 0.13 77.9 ± 2.3 manage the data generated is of critical importance in order to
Indicated 0.65 - 1.1 1734 ± 87 2.84 ± 0.17 174.0 ± 14.0 implement the suggested approaches.
Inferred >1.1 103 ± 9.4 2.92 ± 0.16 10.5 ± 1.9
A LINK TO PRODUCTION SCHEDULING AND
SOME RELATED ASPECTS
important difference from traditional classification methods in
The approaches and examples discussed in the preceding sections
this example is that the classification is independent of the block
do not consider the potential influence of production scheduling to
average grade.
ore reserve classification. It can be argued that sequencing and the
Please note that a key expectation of resource classification is time of mining adds a new dimension to decisions about reserve
to support mine planning. Parker’s approach (Harry Parker, classification. This is because the criteria for deciding what is an
personal communication) commonly used, (shown also by Lopes acceptable risk in one year may not necessarily be the same as that
et al, 2010, in this volume) suggests that Measured should in for decisions ten years later and the risk associated with mining
support three months production increments and Indicated part of an orebody this year should not necessarily be assessed in
annual production increments. Although at the time of this study
the same way as what is to be mined some years later when
this was not implemented, the approach presented can easily and
information and conditions will be different.
should be implemented, so as to state confidence levels in terms
of production increments used for planning. An extension of the example from the alluvial diamond deposit
presented earlier is used here to illustrate the quantification of
uncertainty in reserves (and thus meeting production targets) –
Selecting thresholds at a copper deposit here the total volume of ore material mined. First, a production
schedule is developed through conventional means for twelve 500
The example of a porphyry copper deposit in Canada further × 500 m2 blocks and considers mining four such blocks per period
illustrates the use of a CCV based classification for resource/ over three production periods. Then, uncertainty assessment is
reserve classification. The example considers the part of the performed from a set of simulated orebody models at the same
deposit within the optimal pit limits (see Leite and block support-scale to model the uncertainty in the volume for
Dimitrakopoulos, 2010, in this volume). For the present study, each production period. Figure 10 shows the ccdfs for volumes of
25 realisations of the copper grades of 20 × 20 × 10 m3 blocks are material mined in each of the three production periods, thus
used. Figure 8 shows tonnage of ore above the 0.3 per cent Cu describing their uncertainty. In addition, Figure 10 shows the risk
cut-off used here plotted against the CCV for the 25 simulated profile of the expected volume for three production periods. If risk
realisations of the copper orebody generated directly at the block is a main criterion in reserve classification, then the reserve in the
size required (eg Benndorf and Dimitrakopoulos, 2007). Similarly first mining period should perhaps be classified in a different
to the previous case study, the figure shows a distinctly increasing category from that in the third period.

220 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 RESOURCE/RESERVE CLASSIFICATION WITH INTEGRATED GEOMETRIC AND LOCAL GRADE VARIABILITY MEASURES
120
100
Tonnage above cut-off (tonnes x 106)
80
60
40
20
0
0 0.2
FIG 8 - Tonnage of ore versus the CCV for 25 realisations of the copper deposit.
FIG 9 - Blocks in a section of a gold load classified using the CCV
classification (left) and using available drilling (right).
The integration and understanding of the interaction between
reserve classification with the mining sequence and production
scheduling are traditionally not considered, although in the past,
conditional simulation has been used to test uncertainty of
production schedules (Ravenscroft 1992; and others). However, it
is logical, as the simple example above suggests, to consider
enhancing the classification approaches. For example, the
uncertainty measures previously described may be used to classify
resources/reserves given a production schedule and consider
different classification criteria for the mining periods so as to
reflect timing.
It is also and perhaps more important to stress that scheduling
can be performed under geological uncertainty and geological risk
discounting can be applied within the scheduling formulations.
This discounting concept is introduced in Dimitrakopoulos and
Ramazan (2004) and used in new stochastic long-term mine
production schedulers (Ramazan and Dimitrakopoulos, 2009; also,
Benndorf and Dimitrakopoulos, 2010, in this volume) leads to a
Advances in Orebody Modelling and Strategic Mine Planning I
0.4 0.6 0.8 1 1.2 1.4
CCV
desired blending of materials within periods that provide overall
prespecified risk profile in a scheduling period, thus optimising the
classification of reserves. It is possible to substantially improve
and provide more realistic assessments and classification of
reserves when schedules and resource uncertainty are jointly
managed.
CONCLUSIONS
Mineral resources and ore reserves are classified as Measured,
Indicated, Inferred and Proven or Probable respectively both for
public reporting and internal company assessments. Resource/
reserve classification is traditionally based on geometric criteria,
such as the number of samples or drill holes within a volume, or
some form of estimation variances. Although transparent and
relatively easy to understand, these criteria ignore local variability
of grades within a deposit and can lead to oversimplification or
misrepresentation of the confidence of estimated grades.
In this paper, an approach integrating both geometric criteria
and in situ grade variability is described and its application
explored. The approach is based on conditional simulation and
risk quantification in five steps: generation of suitable
realisations of the deposit at the appropriate block-support scale,
derivation of conditional cumulative distribution functions of
block grades or attribute selected, utilisation of the Conditional
Coefficient of Variation (CCV) of the grade ccdf within a block
as the uncertainty measure for resource classification of each
block in the orebody model, classification of resources/reserves,
and evaluation and production of risk profiles for estimates.
The practical aspects and intricacies of the proposed approach
are explored in three applications at an alluvial diamond deposit,
an epithermal gold lode and a porphyry copper deposit. An
interesting observation, common to these applications, is that there
are natural groupings of the CCV of mining blocks that allow for
rational categorisation of resources or reserves in a transparent and
technically defensible way. With this tool, along with existing
practices, such as examining cross-sections of the deposit, to
ensure the definitions in the National Reporting Codes are met. As
a result, reporting of resource/reserve classification and may
support better investment decision making.
Spectrum Series Volume 17 221
R DIMITRAKOPOULOS, M GODOY and C L CHOU
FIG 10 - Conditional cumulative distribution functions describing the uncertainty on the volume of mineralised materials for the production
periods (left) and risk profile on the expected volume for three production periods (right). Variances plotted are not cumulative.
A key improvement to this approach could be linking it to
mine production scheduling, first, so that classifications account
for production units over time – what is to be mined in ten years
should not perhaps be classified as what is to be mined next year
and secondly, to consider control of the classification of
resources/reserves through risk blending and deferment, a
capability that stochastic schedulers can offer.
ACKNOWLEDGEMENTS
Thanks are in order to the industry members of the COSMO
Laboratory: AngloGold Ashanti, Barrick, BHP Billiton, De
Beers, Newmont, Vale and Vale Inco, as well as NSERC
(Discovery Grant 239019), the Canada Research Chairs Program
and CFI are gratefully acknowledged.
REFERENCES
Arik, A, 2002. Comparison of resource classification methodologies with
a new approach, in Proceedings Applications of Computers and
Operations research in Mineral Industry (APCOM) 2002.
Audet, M and Ross, A F, 2007. Koniambo lateritic Ni-Co deposits, New
Caledonia – A case study from geological modelling to mineral
resource classification, in Orebody Modelling and Strategic Mine
Planning, second edition (ed: R Dimitrakopoulos), pp 235-244 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Benndorf, J and Dimitrakopoulos, R, 2007. New efficient methods for
conditional simulation of large orebodies, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 61-68 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Benndorf, J and Dimitrakopoulos, R, 2010. Stochastic long-term
production scheduling of iron ore deposits: Integrating joint
multi-element geological uncertainty, in Advances in Orebody
Modelling and Strategic Mine Planning I (ed: R Dimitrakopoulos),
pp 151-158 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Canadian Securities Commissions (CSA), 2005. National instrument
43-101: Standards for disclosure for mineral projects.
David, M, 1977. Geostatistical Ore Reserve Estimation, p 364 (Elsevier:
The Netherlands).
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
Reserve Estimation, p 217 (Elsevier: The Netherlands).
David, M and Dagbert, M, 1977. Predicting vanishing tons before
production starts, or small blocks are not good for planning in
porphyry deposits, paper presented to annual meeting of the American
Institute of Mining, Metalurgy and Petroleum Engineering, Atlanta.
Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty based production
scheduling in open pit mining, SME Transactions, 316:106-112.
Dohm, C, 2005. Quantifiable mineral resource classification – A logical
approach, in Geostatistics Banff 2004 (eds: O Leuangthong and
C Deutsch), 1:333-342 (Springer: Dordrecht).
Duggan, S and Dimitrakopoulos, R, 2005. Application of conditional
simulation to quantify uncertainty and to classify a diamond
deflation deposit, in Geostatistics Banff 2004 (eds: O Leuangthong
and C Deutsch), 2:419-428 (Springer: Dordrecht).
222 Spectrum Series Volume 17
Froidevaux, R, Roscoe, W E and Valiant, R I, 1986. Estimating and
classifying gold reserves at Page-Williams C-zone: A case study in
non parametric geostatistics, in Ore Reserve Estimation: Methods,
Models and Reality (eds: M David, R Froidevaux, A J Sinclair and
M Vallée), pp 280-300 (Canadian Institute of Mining: Montreal).
Giroux, G H, Sinclair, A J and Miller, J H L, 1986. Production quality
control experiments, in Ore Reserve Estimation: Methods, Models
and Reality (eds: M David, R Froidevaux, A M Sinclair and
M Vallée), pp 238-260 (Canadian Institute of Mining: Montreal).
Joint Ore Reserves Committee (JORC), 2004. Australasian Code for
Reporting of Exploration Results, Mineral Resources and Ore
reserves (JORC Code).
Journel, A G and Kyriakidis, P C, 2004. Evaluation of Mineral Reserves
– A Simulation Approach, 232 p (Oxford University Press: New York).
Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under
metal uncertainty – Application of stochastic mathematical
programming at an open pit copper mine and comparison to
conventional scheduling, in Advances in Orebody Modelling and
Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Lopes, J A, Rosas, C F, Fernandes, J B and Vanzela, G A, 2010.
A geology-based joint conditional simulation of a lateritic nickel
deposit – Quantification of risk in grade tonnage curves and resource
categorisation, in Advances in Orebody Modelling and Strategic
Mine Planning I (ed: R Dimitrakopoulos), pp 205-214 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Murphy, M, Parker, H, Ross, A and Audet, M, 2004. Ore-thickness and
nickel grade resource confidence at the Koniambo nickel laterite
deposit in New Caledonia – A conditional simulation voyage of
discovery, in Geostatistics Banff 2004 (eds: O Leuangthong and
C Deutsch) (Springer: Dordrecht).
Parker, H. Personal communication.
Ramazan, S and Dimitrakopoulos, R, 2009. Production scheduling with
uncertain supply: A new solution to the open pit mining problem,
European Journal of Operations Research (submitted) and COSMO
Report 2007, pp 257-294.
Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
simulation, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 101:A104-A108.
Sabourin, R L, 1983. Application of Geostatistical method to
quantitatively define various categories of resources, in Geostatistics
for Natural Resources Characterisation Part 1, series C:
Mathematical and Physical Sciences (ed: G Verly) 122:201-216.
Sinclair, A J and Blackwell, G H, 2002. Applied Mineral Inventory
Estimation, 381 p (Cambridge University Press: New York).
South African Mineral Resource Committee (SAMREC), 2000. South
African reporting of mineral resources and mineral reserve, Southern
African Institute of Mining and Metallurgy.
United States Securities and Exchange Commission (SEC), 2006.
Industry Guide 7.
Wawruch, T, Jorge, F and Betzhold, J F, 2005. Mineral Resource
classification through conditional simulation, in Geostatistics Banff
2004 (eds: O Leuangthong and C Deutsch), 1:479-489 (Springer:
Dordrecht).
Advances in Orebody Modelling and Strategic Mine Planning I

Reducing the Geological Risk in Mining an Orebody by Using
Borehole Radar for Strategic Mine Planning
C Kemp1, P Du Pisani2, A Bray3 and G Chitiyo4
ABSTRACT
Borehole radar is a proven geophysical technology that can be used to
reduce the geological risk in extracting an orebody and therefore improve
mine planning. This paper will present a case study, where borehole radar
was used within the mining cycle to map out orebody blocks to assist with
the mine design months ahead of mining. Refined equipment and
procedures now enable a slim-line borehole radar tool to be deployed on
the end of drill rods, making it easier and faster to deploy in an
underground mine with minimal impact on the mining process. The case
study is from a South African platinum mine, where borehole radar was
used successfully in identifying small-scale disruptions in the platinum reef
that may influence the way in which it will be mined. Slim-line borehole
radar tools are deployed in applications where boreholes are drilled parallel
or subparallel to the orebodies in order to delineate them and any disruptive
structures on them, which change the normal orebody location, thickness
and sometimes composition. Advance knowledge of how the orebodies are
displaced impacts on how the orebody is mined. Mine planning can then be
adapted to mine the orebody more economically and potentially hazardous
situations can be negated.
INTRODUCTION
Globally, the mining industry and particularly Anglo Platinum are
adapting to rapidly rising labour costs and improvements in
technology. It is moving from a tradition in which small, flexible
airleg mining crews advance a large number of ~30 m wide
metre-high panels about 17 m per month, through concentrated
capital intensive trackless mining operations advancing at ~50 m
per month on a front five times wider, to experimental continuous
mining operations that are projected to achieve advance rates of
over 100 m per month. Negotiating geological discontinuities
interferes with production to a degree that generally rises with the
level of mechanisation and the rate of face advance. For example,
in platinum reef mining, potholes cause the most frequent
disruptions. A faster advance rate increases the frequency of
exposed potholes – the greater the capital concentration, the more
serious the impact upon ore production. The more unexpected the
pothole, the lower the stockpiles and thus the ripple of feedstock
that enters the mill is greater.
Observations such as these have stimulated recent geophysical
reconnaissance efforts across a number of different commodities.
Three-dimensional surface seismic surveys costing around
AU$250 000 per square kilometre can map faults of several
metres of throw, and most shallow potholes greater than 50 m
wide in plan on those parts of platinum reef that are neither
overshadowed by old workings on an overlying platinum reef,
nor vulnerable to screening by seismic reverberation between
horizons that lie either above or near the platinum reef. This
technique is useful on a broader exploration scale, but more
1. Business Development Manager and Senior Geophysicist, Geomole
Pty Ltd, Australian Technology Park, International Business Centre,
Suite 108, 2 Cornwallis Street, Eveleigh NSW 2015.
Email: carina@geomole.com
2. Manager, In-Mine Geophysics, Anglo Technical Division,
Geosciences Research Group, Office 553, 45 Main Street,
Marshalltown Johannesburg 2107, South Africa.
3. SAusIMM, PhD Student, The School of Geosciences, Madsen
Building (F09), The University of Sydney, Sydney NSW 2006.
4. Chief Geologist, Anglo Platinum, Rustenburg Section, South Africa.
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
detailed information is needed to negotiate geological
discontinuities on the panel scale. The work described in this
paper shows how mining operations can be supported by tactical
underground Borehole Radar (BHR) surveys, and puts forward
the proposition that BHR can be used to better manage many of
the risks that are inherent in concentrating production on a small
number of high output mechanised panels.
BOREHOLE RADAR FOR STRATEGIC MINE
PLANNING OF PLANAR OREBODIES
BHR is an electromagnetic subsurface imaging technique
designed for imaging or detecting discontinuities in restive hard
rock formations. Radar works best when the tool is deployed in a
resistive host rock and used to image a conductive orebody.
Reflections are normally caused by sharp changes in the
electrical properties of the rock, and can yield information on
geological change, faults/fractures, voids, dykes, orebody
geometry, etc. On the other hand, conductive or highly fractured
rock formations increase the attenuation (scattering) of
electromagnetic energy, making BHR less effective.
Instrumentation
A BHR tool consists of a transmitter and receiver, which can be
combined in one antenna, resulting in a ‘single-stick’ probe, or as
two separate antennas in two separate probes, usually separated by
a fibre optic spacer. Both antenna configurations have their
advantages and disadvantages, however due to logistical and space
constraints in underground hard rock mining, usually the
‘single-stick’ probes are used. The current BHR design has come a
long way in the last couple of years. Improving from probes that
were connected by fibre optics to a digital acquisition system at
the borehole collar (Vogt, 2006), to the single-stick radars with
onboard memory (Bray et al, 2007) enabling surveying to be
conducted on the drill rods directly after the boreholes are drilled,
in the same manner that borehole deviation surveys are conducted.
The latter technique now enables BHR to be surveyed in all
underground boreholes, where previously broken ground and
difficult end of hole pulley deployment methods hindered its
success. The slim-line BHR tool has been designed specifically for
underground mining. Its diameter enables it to fit into exploratory
geological or cover drill holes with diameters smaller than 48 mm.
The BHR tool is broadband, with a centre frequency of 50 MHz.
This equates to a range in most hard-rock environments of at least
30 m from the borehole, but up to 80 m in some rock types. The
dielectric properties of the host-rock determine the range that
BHR can achieve. Slim-line BHR tools are omni-directional.
Reflection data is recorded from signals received from 360°
surrounding the boreholes. While a lot of research is being
conducted in directional radar technology, it is difficult to fit an
efficient directional radar antenna in a probe that will fit in
underground mining drill holes, eg less than 48 mm diameter.
Borehole radar survey design
The borehole design and layout is absolutely critical in successful
radar implementation. For thin reef or vein style deposits, BHR is
generally conducted in a series of regularly spaced single-borehole
reflection surveys, however, the borehole design depends greatly
Spectrum Series Volume 17 223
C KEMP et al
on the type of orebody and the geological discontinuities
associated with it. Here we will look at a BHR survey design for
thin reef or vein style deposits.
Survey design is conditioned by limitations of site access and
possible parallel layering due to bedding or joint planes within
the host rock. Consider, in Figure 1, the descent of a BHR
through a succession of bedding planes, the UG2 and UG1.
Passage through each plane in turn generates an inverted ‘V’
pattern in the time-section of a trace gather. Echoes from off-axis
diffractors appear as inverted hyperbolas. Similar hyperbolas
appear on surface seismic profiles and on raw Synthetic Aperture
Radar (SAR) trace gathers. The BHR-SAR analogy is highly
useful, but there are very significant differences between the two.
One lies in the fact that targets of interest in BHR may lie on
many planes. In SAR, most targets lie on the ground. Another is
that directional antennas allow SAR systems to eliminate ground
echoes or nadir reflections and to concentrate on imaging the
backscatter or sideswipe from off-axis targets. In hard-rock BHR,
it is nearly impossible to suppress ground echoes or nadir
reflections, because efficient directional VHF-UHF antennas
cannot fit into narrow (47 mm) boreholes. BHR time sections
tend to be dominated by specular reflections from areas that
straddle nadir lines. Nadir lines are the loci of reflection points
on a reflecting plane. The normals to the surface at these points
collectively define sagittal planes.
If transmitter and receiver are at identical heights above a plane,
the sagittal plane will pass through the bisector. If there are many
parallel boreholes, BHR surveys may paint near-parallel nadir
lines on the bedding planes. The probability P of intersecting an
elliptical pothole or similar disruption, semi-axes a and b, with
(A)
FIG 1 - (A) Example borehole radar layout and (B) resulting borehole radar time section.
(A)
FIG 2 - (A) A pothole within the first Fresnel zone surrounding the nadir line will influence the first arrive, and (B) a small elliptical
pothole lies on a surface generated from a framework of nadir lines.
224 Spectrum Series Volume 17
one of a parallel set of nadir lines is (Davis, 1986) directly
proportional to the ratio of the pothole’s periphery and the nadir
line spacing D:
2( a 2 + b 2 )
P=
D2
The probability of detecting a pothole during a BHR profile of
a borehole is of course higher, for on the one hand many potholes
have deep roots and generate sideswipe, and on the other hand,
even if the pothole does not lie on a nadir line, it can and will
influence the first arrival if it lies within the first Fresnel zone,
shown in Figure 2a. Nadir lines form a framework that may be
used to reconstruct a 3D model of a buried surface, Figure 2b.
Therefore the borehole design for BHR should consider the size
of the object that is required to be detected and the range to the
target plane.
CASE STUDY – PLATINUM REEF MINING
Overview
The 2060 Ma old Bushveld Complex is a massive layered igneous
intrusion, with outcrop extremities of ~450 km east-west and
~300 km north-south. It has variously been attributed to the
subduction of a mid-oceanic spreading centre, a multiple meteorite
impact, and a mantle plume. It can be compared to an enormous,
irregularly-shaped saucer with its centre deeply buried, but its rim
exposed. The saucer is made up of distinct thin layers. Two of
these layers, the Merensky and the underlying UG2 chromitite
(B)
(B)
Advances in Orebody Modelling and Strategic Mine Planning I
 REDUCING THE GEOLOGICAL RISK IN MINING AN OREBODY BY USING BOREHOLE RADAR FOR STRATEGIC MINE PLANNING
reef, are members of a well layered ultramafic to mafic succession
known as the Rustenburg Layered Suite. These two reefs contain
well over half of the world’s platinum reserves (Viljoen and
Schurmann, 1998; Cawthorn, 1999). The reefs are generally
shallow dipping tabular orebodies. They have been tracked using
exploratory drilling and seismic methods over kilometres within
the Bushveld complex (Eales et al, 1993). On a mining scale, the
reefs can be disrupted by various smaller-scale features such as
dykes, faults, slumps (called potholes), rolls and Iron-Rich
Ultramafic Pegmatites (IRUPs). Predicting the location of these
disruptions prior to mining affects resource calculation, mine
planning, the extraction of the orebody, as well as mine safety (Du
Pisani et al, 2008).
BHR has been used successfully in many of South Africa’s
underground platinum mines to delineate both of the the Merensky
and UG2 platinum reefs (Van Schoor, Du Pisani and Vogt, 2006;
De Vries and Du Pisani, 2007). Here we will consider potholes
and rolls that disrupt the reef. According to Lomberg et al (1999),
potholes are more disruptive to layering and mining than rolls.
Mining operations near potholes are disrupted by the sudden dip in
the strata, increasing the number of joints and talc-filled fractures,
cross cutting pegmatite veins and a weakening of the roof. The
reef thins gradually on approach to a shallow conformable
pothole. More rapid reef thinning indicates an approach to a deep
unconformable pothole.
At the edge the upper leader chromitite suddenly
breaks through the lower leader chromitite
layers and then, over two to 5 m, cuts sharply
into the underlying feldspathic pyroxenite …
sometimes … so sharply that it cuts back
underneath the pothole reef (after Lomberg et al,
1999).
Potholes range from one to 500 m in diameter, drop from one to
100 m into the footwall and can have sloping, steep, or even
overhanging margins. They can appear as simple washouts, as
slumps, or as deeply sunken cones, the sides of which may be
decorated by shards of UG2 chromitite and into which are sucked
substantial amounts of the hanging wall. The layers above a deep
pothole are frequently deformed, sometimes plastically. Brittle
failures above some potholes fault the hanging wall and hamper
efforts to mine into or through them.
Borehole radar survey design
BHR test surveys were conducted in 2003 at Anglo Platinum’s
Bleskop shaft that is close to Rustenburg (South Africa) in order
to detect disruptions in the UG2 reef, demonstrate that BHR can
be used to optimise mine layouts and to test the viability of
applying BHR in underground platinum mines. The locations of
boreholes drilled for the application of BHR are limited by the
access available from mining excavations. Platinum reefs in
South Africa are traditionally mined using a conventional breast
stoping mining method. Radar boreholes are usually drilled
below the reef in the footwall, either from haulages or crosscuts
and are oriented parallel or subparallel to the reef plane – see
Figure 3. The area covered by radar boreholes is usually the size
of a planned stope, around 200 × 200 m. This enables mapping
the reef topography prior to mining.
Access to the UG2 at Bleskop was established in the
mid-nineties, through a network of strike haulages that were
driven ten to 20 m below the reef. Four 180 m long AXT
boreholes, numbered 6, 7, 8 and 9 in the perspective view in
Figure 4, were drilled west from stations spaced at ~35 m in a
north-south stub-end that ran out ~10 m beneath a moderately
large pothole in the UG2 from an east-west strike haulage. A fifth
~200 m long tie borehole, Number 10, was drilled south from the
haulage. The holes cover a mining block spanning approximately
20 000 m2. They were rigged with pulleys, surveyed accurately
Advances in Orebody Modelling and Strategic Mine Planning I
FIG 3 - Schematic of borehole radar survey design as seen in
profile (after Du Pisani et al, 2008).
and then shot using BHRs that moved up each hole in turn at
10 m/min.
Borehole radar results and interpretation
The five panels in Figure 5 show the resulting trace gathers. It is
difficult to identify the different reflections in the data from the
Borehole Radar data alone. This is difficult because slimline
BHRs have toroidal radiation patterns and therefore collect data
from 360° around the drillhole. Distinguishing between arrivals
from planes above the borehole and from those below is
accomplished by modelling. The modelling procedure is
illustrated in Figure 6, where four model boreholes track east-west
and a model tie hole tracks north-south. The inclination of the
model mirror plane above the boreholes can be adjusted, as can
the plane’s elevation. On the upper-right of Figure 6, the mirror
appears as a horizontal line, while the borehole appears as a curve.
If the mirror plane’s inclination matches that of the bedding
planes, then as its elevation rises, the mirror move-out line will
successively match the dashed arrivals first from a model footwall
UG1 and then from a model hanging wall UG2.
Figure 7 illustrates the practical application of mirror plane
synthesis. There it is shown in close-up time sections from the
three nearly parallel boreholes Number 06, Number 07 and
Number 08. These were drilled westwards from stations in a
horizontal north-south cross-cut. Their collars were spaced at 35 m
and 25 m. Smooth red curves, superimposed on the time sections,
show the model mirror-plane moveouts at elevations of 778.5,
805.5, 808.8, 812.8 and 818.3 (measured normal to bedding). The
lowest synthetic, at 27.5 m beneath the next – the Bleskop Marker
(BkM) – tracks echoes from the UG1. Echoes from the UG2 (at
808.8 m nominal) indicate a separation of 3.3 m from the BkM.
The echo 4 m higher (at 812.4 m nominal) is linked to a group of
beds known as the triplets. And finally, the top echoes, tracked at
818.3 m, are associated with the Pyroxenite-Anorthosite (PxA)
boundary in the UG2’s hanging wall. Each of the sections can be
examined in detail and converted into a section interpretation. The
section interpretation for borehole Number 8 is shown in Figure 8.
It shows the UG2 breaking away from the normal stratigraphy at
the beginning and end of the hole. Shading to the right indicates
that certainty drops when the borehole, shown in red, ascends into
a perturbed and thinned section of the UG2 reef. The pothole on
the left is approximately 20 m deep and the ‘mine-able corridor’ is
50 m wide. Slumping, reef thinning and shearing mark the western
boundary of this corridor. From this section it is easy to identify
that the centre of the section would be the only suitable place for
undisrupted mining of the UG2 platinum reef.
Spectrum Series Volume 17 225
C KEMP et al

FIG 4 - Perspective view of the radar boreholes drilled at Bleskop.

FIG 5 - Borehole radar gathers, band-pass filtered and gain-swept. Time runs vertically downwards. Distance from hole collars runs
horizontally across the sections. Curved lines track mirror reflections from planes 37.8 m apart.

226 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 REDUCING THE GEOLOGICAL RISK IN MINING AN OREBODY BY USING BOREHOLE RADAR FOR STRATEGIC MINE PLANNING

FIG 6 - Identification of arrivals from specific bedding planes using modelling. Though individual layers may differ locally in thickness, the
stratigraphy, summarised by colours marked both on the bedding plane normal and on the sagittal plane upper right, is known from
geological logs.

FIG 7 - A practical application of mirror plane synthesis: matching time sections from boreholes #06, #07 and #08 to a 3D plane-parallel
layered model of the UG2 reef and its surrounding host rocks. Thin red lines indicate layered stratigraphy. Dotted lines indicate deviations
from this layered stratigraphy.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 227
C KEMP et al

FIG 8 - Section interpretation of the sagittal plane of Borehole Number 08, based on analyses of the sections presented in Figure 5
and Figure 7.

Strata that are healthy from a mechanised mining point of view
usually match layer cake time-of-flight models very well.
Deviations from this layer cake stratigraphy indicate areas of
concern for mining, as disruptions such as potholes have disturbed
the reef in these locations. The resulting interpretation of the
stratigraphy for this mining area is illustrated in Figure 8.
Evidence of disturbance comes from the 2D radar time sections.
The identification of 3D shape comes from interpolating between
arrivals. Recovery of sagittal planes which are consistent with the
interpolated 3D surface can be achieved by iteration. Figure 9 is an
isometric view of the potholed reef, Figure 10 shows the
geological interpretation in plan and Figure 11 shows BHR
derived contours both of the reef and the hanging wall PxA.
Discussion and impact on mining
The results of the BHR indicated various areas of concern for
mining. However, these BHR trials at Bleskop were conducted
when the technique was still considered to be experimental.
Therefore the BHR results were not received in time or with

FIG 9 - Three-dimensional BHR image of a large spoon-shaped UG2 pothole.

228 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 REDUCING THE GEOLOGICAL RISK IN MINING AN OREBODY BY USING BOREHOLE RADAR FOR STRATEGIC MINE PLANNING

FIG 10 - Working interpretation of the surfaces contoured in Figure 9.

FIG 11 - Contours at 2 m intervals link the hanging wall pyroxenite – anorthosite (upper) to the underlying UG2 reef in a potholed panel at
Bleskop Platinum Mine. A large volume of hanging wall material appears to have been drawn laterally into the pothole upper right.

enough weight to influence the mine design and avoidable mining TABLE 1
and development took place. This trial provided a good test case to Financial impact summarised (after Du Pisani et al, 2008).
show how much could have been saved on unnecessary mining
and development when using BHR. Subsequent mining of this Item Extent Tonnes Cost
mining block (green panels in Figure 12) showed that the Avoidable mining 4300 m2 14 400 US$1.03 M
conclusions drawn from the BHR results were true and that the
Avoidable development 338 m US$230 k
large anomaly identified on the western side of the block was a
significant reef roll. Figure 13 illustrates the unnecessary mining Total US$1.26 M
and development that took place in this mining area at Bleskop.
The costs that could have been saved within the Bleskop African platinum mines. It should also be noted that not all
pothole, if the BHR results had been used, are summarised in panels lend themselves readily to high-levels of mechanisation.
Table 1 (after Du Pisani, 2008). All costs used are based on Many contain unsuspected faults that may delay or halt
November 2007 Platinum-Group Metals (PGM) prices as well as production. BHR surveys can add significantly to the value of an
generalised costs for mine development and drilling at South orebody not only by enabling optimum equipment to be selected

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 229
C KEMP et al

FIG 12 - Geological features encountered during subsequent mining of Bleskop mining block (after du Pisani et al, 2008).

FIG 13 - Unnecessary mining and development that took place within the Bleskop pothole (after Du Pisani et al, 2008).

for a given task, but also by allowing the effort put into the
development of different faces to be synchronised so that they do
not all go down together. BHR surveying adds to predevelopment
costs. A kilometre of AXT drilling in the footwall takes about
one hundred pneumatic crew shifts and costs about AU$50 000.
Boreholes are 50 to 100 times as expensive as BHR. This means
that the BHR surveying of a 200 m × 200 m stope from the
footwall that involves one kilometre of boreholes would cost less
than ten to 20 m of unnecessary development.
SUMMARY AND CONCLUSIONS
The case study presented in this paper shows that BHR, if applied
correctly, can significantly contribute to lowering the costs of
operating a platinum mine. Strategically conducting BHR surveys

230 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 REDUCING THE GEOLOGICAL RISK IN MINING AN OREBODY BY USING BOREHOLE RADAR FOR STRATEGIC MINE PLANNING
both prior to and within the mining cycle improves the geological
knowledge about the orebody. BHR surveys enable strategic mine
planning for the appropriate mine design and the efficient
deployment of mining teams. But most of all, BHR can identify
additional resources in structurally complex areas of the mine
and sterilise areas that are too disrupted to mine. At Anglo
Platinum, BHR has since moved beyond the realm of trial
surveys and is being routinely applied as a geological mapping
tool at various mines.
ACKNOWLEDGEMENTS
The authors thank Anglo Platinum for allowing these results to be
published. In particular, Theo Pegram, Gordon Chunnett and
Marshall Patterson are thanked for their continued support of
borehole radar as a delineation tool for platinum reefs. Tony
Redman (Anglo American), Charles Pretorius (ATD) and Alan
King (ATD) are acknowledged for their support, as are all the
Rustenburg geologists, borehole radar surveyors and drilling
contractors who have made these borehole radar surveys possible.
The authors would like to acknowledge staff members and miners
of RPM Rustenburg Section for their vision and support of the
implementation of geophysical programmes. The professional
enthusiasm of Francois Vos, Vasek Novi, Kabelo Thlapi and Luke
Zindi made this study both possible and enjoyable. We thank the
drillers of Rosond and all of the miners at Brakspruit, Bleskop,
Waterval and Boschfontein for their competence and invariably
kind assistance. The authors gratefully acknowledge the scientific
and technical contributions of Iain Mason, Jonathan Hargreaves,
Binzhong Zhou, Lang Hames, Brian Woods, Wessel van Brakel
and Paul van der Merwe.
REFERENCES
Bray, A, Sindle, T, Mason, I M, Palmer, K, Cloete, J, Steenkamp, J, Du
Pisani, P and Trofimczyk, K, 2007. Using slimline borehole radar
from cover holes, in Proceedings Tenth SAGA Biennial Technical
Meeting and Exhibition (South African Geophysical Association).
Advances in Orebody Modelling and Strategic Mine Planning I
Cawthorn, R G, 1999. The platinum and palladium resources of the
Bushveld Complex, South African Journal of Science, 95:481-489.
Davis, J C, 1986. Statistics and Data Analysis in Geology, second edition,
p 290 (John Wiley: New York).
De Vries, P and Du Pisani, P, 2005. Borehole radar delineation of the
UG2 reef at Modikwa Mine, in Proceedings Ninth SAGA Biennial
Technical Meeting and Exhibition (South African Geophysical
Association).
Du Pisani, P, Coomber, S, Chitiyo, G, Daniso, V, Mampa, S and Mason, I,
2008. Using borehole radar to detect disruptions in platinum reefs in
South Africa – The financial implication, in Proceedings 12th
International Conference on Ground Penetrating Radar, 16-19 June,
Birmingham, UK.
Du Pisani, P, Mason, I M and Daniso, L N, 2007. Borehole radar
delineation of the Brakspruit regional pothole at Anglo Platinum’s
Rustenburg Section, in Proceedings Tenth SAGA Biennial Technical
Meeting and Exhibition, (South African Geophysical Association).
Eales, H V, Botha, W J, Hattingh, P J, De Klerk, W J, Maier, W D and
Odgers, A T R, 1993. The mafic rocks of the Bushveld Complex: A
review of emplacement and crystallisation history and
mineralisation, in light of recent data, Journal of African Earth
Science, 16:121-142.
Lomberg, K G, Martin, E S, Patterson, M A and Venter, J E, 1999. The
morphology of potholes in the UG2 Chromitite Layer and Merensky
Reef (pothole reef facies) at Union Section, Rustenburg Platinum
Mines, South African Journal of Geology, 102(3):209-220.
Van Schoor, M, Du Pisani, P and Vogt, D, 2006. High-resolution,
short-range, in-mine geophysical techniques for the delineation of
South African orebodies, South African Journal of Science,
102:355-360.
Viljoen, M J and Schürmann, L W, 1998. The Mineral Resources of South
Africa (Council for Geoscience: Pretoria).
Vogt, D, 2006. A borehole radar system for South African gold and
platinum mines, South African Journal of Geology, 109:521-528.
Spectrum Series Volume 17 231
 HOME

Strategic Optimisation of a Vertical Hoisting Shaft in the Callie

Underground Mine
M G Volz1, M Brazil2 and D A Thomas3

ABSTRACT
The Callie underground mine, located in the Tanami Desert in the
Northern Territory, includes two parallel declines accessing a large
orebody extending some two kilometres below the surface. One of several
ideas considered in strategic mine planning is to incorporate a vertical
hoisting shaft and an orepass as an alternative to trucking material to the
surface along the declines. In this work, we use network optimisation
techniques to investigate the feasibility of the proposed system, and to
mathematically determine the optimum positions and geometry of the
shaft, orepass and surrounding infrastructure. We propose a modelling
procedure taking aspects from a mathematical problem, called the Fermat-
Weber problem, which asks for a point minimising the sum of weighted
distances to a given set of points. We describe the implementation of the
procedure into a computer program for solving the problem iteratively, and
present results over a range of infrastructure and haulage costs, decline
gradients and life-of-mine (LOM) schedules.

INTRODUCTION
Located in the Tanami Desert in the Northern Territory,
approximately 550 km north-west of Alice Springs, the Tanami
operations comprise a processing facility, several open pit mines, FIG 1 - Typical cross-section through the orebody (looking west).
the Tanami Mill and the Callie Underground Mine. Gold was
discovered at Tanami in 1900 and modern mining began in 1983
following an agreement with traditional landowners. Initial
production came from the open pit mines at the Granites, while
current production comes from the high-grade Callie underground
mine at Dead Bullock Soak. The Callie underground mine
services a large orebody running approximately in an east-west
direction, plunging into the ground at an angle of about 45°
towards the east (Figure 1). The orebody is divided into two major
veins. At the time of the study, the orebody had been accessed by
a single decline, called the Callie decline, with material having
been mined to a depth of about 1000 m below the surface
(Figure 2). Rock is extracted from the orebody in primary stopes,
which are replaced with fill material to allow adjacent secondary
stopes to be subsequently extracted. Once loaded onto trucks, ore
is hauled along cross-cuts (horizontal tunnels) to the decline.
Levels are at 40 m vertical intervals. In addition to the planned
extension of the primary Callie decline, a secondary decline,
called the Wilson Drill Decline (WDD), is to branch out from the
FIG 2 - The Callie underground mine (looking north), including the
Callie decline about 950 m below the surface (Figure 2). Both
existing Callie decline (white), its planned extension (red/blue) and
declines are to have fixed gradients. Levels servicing the WDD are
the proposed Wilson drill decline (pink).
also at 40 m vertical intervals, however they are offset from the
Callie levels. The Callie decline services the Wilson shoot, while
the WDD is to service a second shoot called the Federation shoot. hoisted to the surface via the shaft. This method can provide
One of several ideas considered in strategic mine planning is significant reductions in operating costs, although it requires a
to incorporate a vertical hoisting shaft and an orepass as an large capital cost associated with a hoisting shaft. In addition to
alternative to trucking material to the surface along the declines. the shaft, it was proposed to include an orepass into the mine. An
Using this system, ore is hauled to a common tipping level, orepass is a near-vertical chute down which ore from upper
called the haul level, where it is crushed, loaded into a skip and levels is dropped to the haul level (which is three levels above the
shaft base), and transported to the shaft in one of several ways,
for example:
1. Analyst, TSG Consulting, Level 11, 350 Collins Street, Melbourne
Vic 3000. Email: marcus.volz@tsgconsulting.com.au
• Ore is loaded into a truck by a load-haul-dump vehicle at the
bottom of the orepass. It is then trucked from the bottom of
2. Senior Lecturer, Department of Electrical and Electronic the orepass to the shaft.
Engineering, The University of Melbourne, Parkville Vic 3010.
Email: brazil@unimelb.edu.au • A load-haul-dump vehicle trams ore directly from the base of
the orepass to the shaft.
3. Professor and Head, Department of Mechanical and Manufacturing
Engineering, The University of Melbourne, Parkville Vic 3010. • If the horizontal distance between the orepass and the shaft is
Email: doreen.thomas@unimelb.edu.au greater than say 300 m, a loading chute may be installed at the

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 235
M G VOLZ, M BRAZIL and D A THOMAS
base of the orepass, allowing trucks to be loaded automatically
before transporting ore to the shaft. For operational reasons the
orepass is assumed to be constrained to the northing 9250N,
which is about halfway between the two declines in plan.
The primary goal of this investigation is to mathematically
determine:
• the optimum position (depth and plan coordinates) of the
hoisting shaft;
• the optimum position (top, bottom and plan coordinates) of
the orepass and identification of the levels which access the
orepass; and
• the optimum geometry of the main haulage drive network at
the tipping level and shaft haulage level.
The analysis is undertaken over a range of infrastructure and
haulage costs, decline gradients and life-of-mine schedules.
PROBLEM DATA
The nominal data used for the purposes of this project is
summarised here and includes mine costs, decline gradients,
access points, life-of-mine schedules and no-go zones. Mine
costs are listed in Table 1. A range of values have been provided
for some of the items, so that sensitivity analyses could be
undertaken to examine the effects of these parameters. All other
costs are assumed to be fixed, in the sense that they are invariant
to the geometry of the shaft, orepass and surrounding
infrastructure. At the time this study was undertaken, the Callie
and Wilson Drill declines were designed to about 1400 m below
the surface with gradients 1:8 and 1:7 respectively. It is assumed
that both declines will continue downwards in the same manner
as the current design. An alternative scheme is to have the Callie
decline gradient equal to 1:7 and the WDD gradient set at 1:6.
TABLE 1
Mine costs.
Component Cost
Shaft development $25 000/m, $50 000/m or
$75 000/m
Orepass development $1210/m
Level and haul drive development $3265/m
Haulage up decline and across $0.75/(t.km) or $1.05/(t.km)
levels
Haulage down decline $0.85/(t.km) or $1.20/(t.km)
Orepass fit-out $1 M
Access points are locations where the nominal cross-cuts
intersect the declines. Each access point is designated a nominal
level corresponding to its approximate Reduced Level (RL). The
surface is at approximately 1400 m RL, and access points on the
Callie decline extend from 340 m RL to -660 m RL at 40 m
vertical intervals (26 points), while access points on the WDD
extend from 390 m RL to 70 m RL, also in 40 m vertical
increments (nine points). Hence in total there are 35 access points.
Their associated declines, (X, Y and Z) coordinates and nominal
levels are not provided in this paper. Three life-of-mine production
schedules for the Callie underground mine were proposed. They
are base, probable and best. Details of the three schedules are
provided in Table 2. The predicted tonnage to be accessed from
each of the 35 access points can be determined for the three
schedules. To avoid disruption to the shaft over the life of the
mine, it must avoid impinging on several no-go regions. Firstly,
the shaft must not be too close to the orebody. The no-go region is
modelled as a barrier around the Wilson and Federation orebodies.
Polygons digitised around the boundaries of the orebodies were
expanded 200 m in any direction. Secondly, the shaft is required
236 Spectrum Series Volume 17
TABLE 2
Base, probable and best life-of-mine schedules.
Callie decline Wilson drill decline
Rate Quantity Final Rate Quantity Final
Name
(Mt/a) (Mt) year (Mt/a) (Mt) year
Base 2 10 2015 0.35 1.4 2014
Probable 2.5 17.5 2018 0.5 2.5 2015
Best 3.5 30 2019 0.5 3.5 2017
to avoid faults by at least 50 m. The main haulage drive is
allowed to pass through a fault so long as it does not travel along
(parallel) to it for any great length, say, no more than 15 m at a
time. Thirdly, the preferred area for the shaft collar is in a region
south of the main entrance road.
After a preliminary analysis of the model, it was identified that
the faults and surface infrastructure constraints could be relaxed.
For the remainder of this paper we assume that the only the
orebody standoff constraint is enforced.
PROBLEM FORMULATION AND SOLUTION
PROCEDURE
Mathematical network
Figure 3 shows: (i) a perspective view; and (ii) a plan view of the
shaft, orepass and surrounding infrastructure.
Under this arrangement, ore from the lower levels is trucked
up both declines to a level, L1, from where it is transported to the
base of the shaft via a horizontal tunnel. Ore from the upper
levels is trucked down both declines to another level, L3, from
which it is transported via a horizontal tunnel to the top of the
orepass. It is dropped down the orepass to L1 and transported to
the shaft base. Ore from levels between L1 and L3 on the
secondary decline is either trucked up to L3 or down to L1,
whichever is closest. Ore from levels between L1 and L3 on the
Callie decline also has the option of being trucked to an
intermediate level, L2, which is between the top and bottom of
the orepass. It is transported to a second tipping point to the
orepass at L2 via a horizontal tunnel, dropped down the orepass
to L1 and transported to the shaft base.
Let p1, p2 be points on the Callie decline and WDD respectively
at L1, p3 the point on the main decline at L2 and p4,, p5 points on
the two respective declines at L3. If levels L1, L2 and L3 are
known, the five points are fixed points corresponding to access
points on the two declines. Denote the plan location of the orepass
by x and that of the shaft by p. While the orepass is allowed to be
positioned anywhere on the northing 9250N (which runs between
the two decline centrelines), the hoisting shaft must avoid the
orebody standoff no-go region. The boundary of this orebody
standoff zone is modelled by the straight line l. Clearly, the
optimal position of p will be on l, rather than behind it, since in the
latter case, the length of the tunnel connecting the two points can
be reduced by moving p onto l. The system of tunnels can be
modelled as a mathematical network T with a star topology, where
x is the centre of the star, p1,…, p5 are fixed and p is free to slide
along l.
For a given LOM schedule, the quantities t1,…, t5 of ore (in
tonnes) can be computed for each fixed point p1,…, p5. For
example, the quantity of ore assigned to p1 is the sum of
tonnages from the levels on the Callie decline below L1 and the
levels immediately above L1 on the same decline.
Let d denote the cost per unit length of developing a tunnel,
and h the cost per unit length of hauling a unit quantity of ore
along a level tunnel. Then the sum of development and haulage
costs for the tunnel from p1,…, p5, to x is (d+hti)li, where li is the
length of the tunnel from pi to x.
Advances in Orebody Modelling and Strategic Mine Planning I
 STRATEGIC OPTIMISATION OF A VERTICAL HOISTING SHAFT IN THE CALLIE UNDERGROUND MINE
FIG 3 - (i) Perspective view and (ii) plan view of the shaft, orepass
and surrounding infrastructure.
Thus the associated weight is wi=d+hti. Similarly, the weight
associated with the tunnel from x to p is:
5 by the first subproblem. The optimum network was
w6 = d + h∑ ti ,
i=1
since that tunnel routes all the ore from the orepass base to the
shaft base.
The total cost of the mine (ignoring fixed costs) has the follow-
ing components:
• shaft development, which depends on the shaft depth
(determined by L1);
• decline haulage, which depends on the levels L1, L2 and L3;
and
• surrounding infrastructure development and haulage, which
depend on the levels L1, L2 and L3 and the plan positions of
the orepass and shaft.
Hence, the total cost of the mine can be minimised by
determining an optimal layout of the shaft, orepass and surround-
ing infrastructure.
Advances in Orebody Modelling and Strategic Mine Planning I
The Fermat-Weber problem
The problem of minimising the development and haulage costs
associated with the horizontal tunnels connecting the declines,
orepass and shaft can be solved by projecting everything onto a
horizontal plane (Brimberg, Juel and Schobel, 2002). Then the
problem reduces to positioning x and p so as to minimise the
function:
6
f ( x , p ) = ∑ wi li
i=1
This is an extension of a well known mathematical problem
called the Fermat-Weber problem. The problem asks for a point,
called a Fermat-Weber point, minimising the sum of weighted
distances to a set of given points in space. In this case the orepass
position is the Fermat-Weber point, while the points p1,…, p5 and
p are the fixed points.
An added complexity is that the shaft is not fixed, but is free to
slide along the line l. We can immediately state the following
result. If x=x0 and p=pk minimise the cost function f(x,p), then
the line segment between x0 and pk is perpendicular to l. Thus,
the position of p is a function of the position of x. This can be
seen by noticing that if x0p is not perpendicular to l, then the
length of x0p can be reduced by moving p along l until it is.
Hence, for given L1, L2 and L3, the optimum network of tunnels
interconnecting the declines, orepass and shaft can be obtained
by solving an extension of the Fermat-Weber problem. This
problem can be solved iteratively using a well known descent
algorithm due to Weiszfeld (1937). For this study, the algorithm
was generalised to account for the fact that the shaft can lie
anywhere on the orebody standoff boundary and is not confined
to a fixed point. We do not give details of this amended algorithm
in this paper.
Iterative solution procedure
Initially, we proposed a simplified procedure for determining the
positions, lengths and orientations of the shaft, orepass and main
haulage drive. By this procedure, the problem was broken down
into two subproblems:
1. The optimum shaft depth was computed based on ‘vertical’
costs – decline haulage, shaft development and orepass
development – using a further extension of the
Fermat-Weber problem, called the gradient-constrained
Fermat-Weber problem (see Brazil, Rubinstein and Volz,
2005).
2. A minimum-cost network interconnecting the declines,
orepass and shaft was constructed at the level determined
computed based on ‘horizontal’ costs – haul drive
development and haulage – using the Fermat-Weber
problem described above.
Although the simplified procedure seems to obtain good
intuitive solutions, it does not guarantee an optimal solution for a
given problem. To see this, suppose the haul level is placed at
100 m RL based on vertical costs. Numerical tests have shown
that moving the haul level to 140 m RL can reduce the cost of the
network of tunnels. If the reduction in horizontal costs outweighs
the increase in vertical costs, then the solution determined by the
simplified procedure is not optimal. Hence, we propose a
rigorous procedure which guarantees an optimal solution for a
given problem. Refer to Figure 3. Placing the haulage drive L1 at
each of the 35 levels (on both declines) from -660 m RL up to
390 m RL, the vertical costs can be computed. If an orepass is
used, it is tested at every level above the haulage drive, and if an
intermediate tipping point is added to the orepass, it is also tested
at every level between the haulage drive and the top of the
Spectrum Series Volume 17 237
M G VOLZ, M BRAZIL and D A THOMAS
orepass. For every possible arrangement of L1, L2 and L3, a
minimum-cost network interconnecting the declines, orepass and
shaft is computed using the amended Fermat-Weber problem.
Once the cost for every mine layout has been computed, the
arrangement giving the lowest cost is selected as the globally
optimal solution.
RESULTS
The mathematical model and algorithm described in the previous
section were implemented into a computer program for
determining the optimum shaft, orepass and surrounding infra-
structure layout for a given set of infrastructure and haulage costs,
decline gradients and life-of-mine schedules. Thirty-six tests
were undertaken for each of three cases – no orepass, one
orepass and one orepass with two tipping points (108 tests in
total).
The gradients and cost parameters were varied across the tests.
Optimum haul levels and costs for the three orepass configurations
are compared in Table 3.
ANALYSIS
In this section we analyse the results of Table 3.
Shaft depth and location
The optimum haul level ranges from 390 m RL to -340 m RL
(1010 m to 1740 m below the surface). A breakdown of
minimum and maximum haul levels is shown in Table 4. The
shaft optimally lies on the orebody standoff boundary between
points (60097, 9043) and (60667, 8900), over 588 m.
Orepass versus no orepass
If an orepass is justified, it is always economical for it to have
two tipping points. For the ‘base’ schedule, an orepass is not
justified, except if the shaft development is $25 000/m, in which
case it is generally more economical to include an orepass with
two tipping points. For the ‘probable’ schedule, an orepass with
two tipping points is justified in all cases except PR08 and PR09.
In these latter cases an orepass is not justified. For the ‘best’
schedule, an orepass with two tipping points is always justified.
The orepass optimally lies on 9250N between 60250E and
60755E, over 505 m.
Effects of parameters
We now discuss the effects of changing various parameters on
the output of the model.
Shaft development
A breakdown of costs, averaged over all 108 tests, was
computed. This breakdown showed that shaft development is by
far the most significant cost component, followed by decline
haulage. This explains why varying the shaft development from
$25 000/m to $50 000/m to $75 000/m has such a significant
impact on the shaft depth and the total cost of the mine. On
average, increasing the shaft development from $25 000/m to
$50 000/m causes the optimum haul level to move up by three
40 m levels. On average, increasing the shaft development from
$50 000/m to $75 000/m causes the optimum haul level to move
up by one 40 m level. Thus, increasing the shaft development
causes a significant decrease in the shaft depth, because the
resulting increase in decline haulage is outweighed by the
decrease in shaft development.
238 Spectrum Series Volume 17
Orepass
The following effects of including an orepass in the model were
observed:
• on average, including an orepass with one tipping point
causes the optimum haul level L1 to move down by three
levels;
• on average, including an orepass with two tipping points
causes the optimum haul level to move down by seven levels;
• including an orepass provides additional savings by reducing
decline haulage; and
• provision of a second tipping point introduces further
savings.
Providing an orepass causes the optimum haul level to become
lower. This is because decline haulage is fixed for levels serviced
by the orepass, ie the cost of decline haulage for these levels is
independent of the shaft depth and is ignored in the model.
Hence the shaft tipping level must be lowered to account for the
unbalanced haulage, shaft and orepass components. Including
additional tipping points causes the shaft depth to increase
further. The more ore is dropped down the orepass, the deeper
the shaft will be.
Gradients
The effect of increasing the Callie and WDD gradients from 1:7
and 1:6 to 1:8 and 1:7 respectively is to increase decline haulage
costs. If decline haulage costs were very large compared to shaft
development, the orepass and shaft tipping levels would tend to
space themselves out over the depth of the mine, so as to
minimise the average haulage distance. Consequently the shaft
depth would increase. In this study, shaft development is very
large compared to decline haulage, and the effect of increasing
gradients has little or no effect.
The following observations were noted when the model was
run with decline gradients set to zero. For the no orepass case,
the optimum haul level is 180 m RL. If an orepass with one
tipping point is used, the optimum haul level is -460 m RL, the
top of the orepass is 180 m RL, upwards haulage is from the
level at -180 m RL. If an orepass with two tipping points is used,
the optimum haul level is -160 m RL, the top of the orepass is
180 m RL (haul up from -20 m RL), and the second tipping point
is at -180 m RL (haul up from -300 RL). It can be verified that,
in all three cases, the haul and orepass levels have positioned
themselves so as to minimise the total haulage cost.
Haulage costs
On average, increasing the decline haulage costs from $0.75/t.km
and $0.85/t.km for upwards and downwards haulage to
$1.05/t.km and $1.20/t.km respectively has the effect of lowering
the optimum haul level by one 40 m level. The reasons are
similar to the reasons for the gradient effects.
One versus two tipping points
The provision of more than two tipping points could potentially
result in further savings, although the additional cost of
developing tunnels from the declines to the orepass must be
considered. Moreover, experiments have indicated that having
two orepasses, one for each decline, reduces haulage and
development costs associated with the haulage drive which, on
average, account for about 12 per cent of the total variable cost
of the mine.
Advances in Orebody Modelling and Strategic Mine Planning I
 STRATEGIC OPTIMISATION OF A VERTICAL HOISTING SHAFT IN THE CALLIE UNDERGROUND MINE

TABLE 3
Summary of results.

Case Schedule Shaft Haulage Haulage Callie WDD No orepass One orepass Orepass with Best option
ID development up down decline gradient two tipping
gradient points
Haul Cost Haul Cost Haul Cost
level ($M) level ($M) level ($M)
BA01 Base 25 000 0.75 0.85 1:8 1:7 158 53.4 19 53 -21 52.6 Two tipping points
BA02 Base 25 000 0.75 0.85 1:7 1:6 158 51.4 19 51.8 -21 51.6 No orepass
BA03 Base 25 000 1.05 1.2 1:8 1:7 99 60.8 -21 58.8 -101 57.7 Two tipping points
BA04 Base 25 000 1.05 1.2 1:7 1:6 99 58.3 -21 57.3 -21 56.5 Two tipping points
BA05 Base 50 000 0.75 0.85 1:8 1:7 299 83.1 219 85.3 219 85.2 No orepass
BA06 Base 50 000 0.75 0.85 1:7 1:6 299 80.3 278 82.8 259 82.8 No orepass
BA07 Base 50 000 1.05 1.2 1:8 1:7 219 92.1 99 92.8 99 92.8 No orepass
BA08 Base 50 000 1.05 1.2 1:7 1:6 219 88.9 99 90.6 99 90.6 No orepass
BA09 Base 75 000 0.75 0.85 1:8 1:7 398 109.2 358 112.6 318 113.2 No orepass
BA10 Base 75 000 0.75 0.85 1:7 1:6 398 105.7 358 109.3 318 110.2 No orepass
BA11 Base 75 000 1.05 1.2 1:8 1:7 299 121 278 123.7 259 123.7 No orepass
BA12 Base 75 000 1.05 1.2 1:7 1:6 398 116.9 278 120 259 120.2 No orepass
PR01 Probable 25 000 0.75 0.85 1:8 1:7 59 68.6 -21 64.6 -101 63 Two tipping points
PR02 Probable 25 000 0.75 0.85 1:7 1:6 99 65.4 -21 62.6 -101 61.5 Two tipping points
PR03 Probable 25 000 1.05 1.2 1:8 1:7 59 81.6 -101 74.5 -101 71.6 Two tipping points
PR04 Probable 25 000 1.05 1.2 1:7 1:6 59 77.2 -21 71.9 -101 69.6 Two tipping points
PR05 Probable 50 000 0.75 0.85 1:8 1:7 158 100.9 19 99.8 -21 99 Two tipping points
PR06 Probable 50 000 0.75 0.85 1:7 1:6 158 97.3 99 97.3 19 97.1 Two tipping points
PR07 Probable 50 000 1.05 1.2 1:8 1:7 99 114.7 -21 110.3 -21 108.8 Two tipping points
PR08 Probable 50 000 1.05 1.2 1:7 1:6 99 110.1 -21 107.5 -21 106.2 Two tipping points
PR09 Probable 75 000 0.75 0.85 1:8 1:7 219 131.1 99 132.7 99 132.6 No orepass
PR10 Probable 75 000 0.75 0.85 1:7 1:6 219 126.9 179 129.5 158 129.4 No orepass
PR11 Probable 75 000 1.05 1.2 1:8 1:7 158 146.5 59 145.1 -21 144.3 Two tipping points
PR12 Probable 75 000 1.05 1.2 1:7 1:6 158 141.4 99 141.3 99 141.2 Two tipping points
BE01 Best 25 000 0.75 0.85 1:8 1:7 -21 100.6 -221 89.2 -261 85.3 Two tipping points
BE02 Best 25 000 0.75 0.85 1:7 1:6 -21 94.3 -181 85.7 -261 82.6 Two tipping points
BE03 Best 25 000 1.05 1.2 1:8 1:7 -61 125.6 -221 107.1 -341 101 Two tipping points
BE04 Best 25 000 1.05 1.2 1:7 1:6 -21 116.8 -221 102.3 -261 97.4 Two tipping points
BE05 Best 50 000 0.75 0.85 1:8 1:7 19 135.7 -141 128.3 -141 126.5 Two tipping points
BE06 Best 50 000 0.75 0.85 1:7 1:6 59 129.1 -101 124.2 -141 122.8 Two tipping points
BE07 Best 50 000 1.05 1.2 1:8 1:7 -21 161.2 -181 147.3 -261 142.8 Two tipping points
BE08 Best 50 000 1.05 1.2 1:7 1:6 19 152.3 -141 141.9 -181 138.9 Two tipping points
BE09 Best 75 000 0.75 0.85 1:8 1:7 77 169.5 -101 166 -61 164.7 Two tipping points
BE10 Best 75 000 0.75 0.85 1:7 1:6 139 162.3 -21 161.1 -21 160.2 Two tipping points
BE11 Best 75 000 1.05 1.2 1:8 1:7 19 195.9 -141 185.9 -181 183.1 Two tipping points
BE12 Best 75 000 1.05 1.2 1:7 1:6 59 186.5 -101 179.9 -141 178.2 Two tipping points

Confinement of orepass to fixed northing TABLE 4

Optimum haul levels.
Confining the orepass to be fixed on the northing between the
two declines has the effect of increasing the main haulage drive Schedule Minimum haul Maximum haul Range
development and haulage costs. The increase, however, is not level level
significant.
Base -60 m RL 390 m RL 450 m
CONCLUSIONS Probable -220 m RL 350 m RL 570 m
Best -340 m 310 m RL 650 m
In this work, we have developed and a network model for the
Callie underground mine. An algorithm was developed to
mathematically determine an optimum location and depth of a algorithm was implemented into a software product for solving
vertical hoisting shaft in the Callie underground mine. The the problem iteratively. Results were analysed over a range of

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 239
M G VOLZ, M BRAZIL and D A THOMAS
infrastructure and haulage costs, decline gradients and life-of
-mine schedules.
ACKNOWLEDGEMENTS
We thank Newmont Australia Limited and the Australian
Research Council for their joint sponsorship of this research via
an ARC Linkage Grant. In particular, we thank Newmont for
suggesting and formulating the Callie shaft location study. A
simplified version of study was originally proposed by Steven
Harvey and the detailed investigation was proposed and
commissioned by Andrew Fox and Ian Suckling. I thank Robert
Parr and Nadine Wetzel for their time and effort collecting and
preparing data and for their constructive feedback. Further
details of this work can be found in the author’s PhD thesis
(Volz, 2008).
240 Spectrum Series Volume 17
REFERENCES
Brazil, M, Rubinstein, J H and Volz, M, 2005. The gradient-constrained
Fermat-Weber problem for underground mine design, in Proceedings
18th National ASOR Conference and Eleventh Australian
Optimisation Day (eds: L Caccetta and V Rehbock), pp 16-23.
Brimberg, J, Juel, H and Schobel, A, 2002. Linear facility location in
three dimensions – Models and solution methods, Operations
Research, 50(6):1050-1057.
Kupitz, Y S and Martini, H, 1997. Geometric aspects of the generalised
Fermat-Torricelli problem, in Intuitive Geometry (eds: I Barany and
K Boroczky), 6:55-127 (Janos Bolyai Mathematical Society).
Volz, M G, 2008. Gradient-constrained flow-dependent networks for
underground mine design, PhD thesis (unpublished), University of
Melbourne, Melbourne.
Weiszfeld, E, 1937. Sur le point pour lequel la somme des distances de n
points donnes est minimum, Tohoku Mathematical Journal,
43:355-386.
Advances in Orebody Modelling and Strategic Mine Planning I

Optimising Value for an Underground Project Configuration
T Elkington1, R Durham2 and P Myers3
ABSTRACT
Strategic mine planning practices for open pit mines have been well
established over many years through processes generally referred to as
‘optimisation’. Such methods for underground mines are not as
advanced, due in part to the added complexity and flexibility of these
operations. This paper presents an approach for maximising the value of
a specified underground project configuration – the case study of a
hypothetical vertical gold deposit demonstrates the substantial value
increases and risk reduction that can be attained by optimising cut-off
grade, rather than staying with the traditional break-even approach. The
ability to optimise cut-off grade spatially is analysed and shown to
increase net present value by 24 per cent and decrease mine life by six
years. The ability to rapidly optimise a single configuration enables
many configurations to be tested to determine the optimal project
strategy with which to progress to a more detailed study phase.
Opportunities for optimisation and value enhancement analysed in this
paper include production rate and price sensitivity.
INTRODUCTION
Strategic mine planning is the process of finding the
configuration that will optimise project objectives. These
objectives may be value based, risk based or based on some other
goal, depending on the company and the situation. A value based
objective, specifically net present value (NPV), is most often
applied. The configuration of an underground operation consists
of a number of key components that can be generally categorised
into geology, mining, processing and sales. The geology
component describes the orebody, while the mining component
describes the mining method, haulage plan, development concept
and mining rate. The processing component includes the method
of processing and the processing rate, while the sales component
considers the nature of the product and how it is sold. These
components each have a series of associated parameters, such as
the orebody geometry and its grade/tonnage relationship, costs,
recoveries and revenues.
The analysis of a project configuration requires the
development of a strategic mine design and plan that matches the
configuration. The mine design and plan is generally
characterised by cut-off grade selection, stope design,
development design, production scheduling, and development
scheduling. Each plan will have associated cash flow and value
measures, which can be used to compare alternatives. A
consistent basis of comparison between project configurations is
essential for informed decision support.
The concept of varying grade to maximise discounted value
was first proposed by Berry (1922); citing the advantages of
mining the high grade portion of an orebody first. Raising the
cut-off grade above the zero profit point was mathematically
quantified by Lane (1964). Lane stated that the optimal cut-off
grade is influenced by the present value of the mine. This
implied that the cut-off decision is dependent on the future mine
plan which was, in turn, dependent on the cut-off decision;
1. MAusIMM, Lecturer, School of Civil and Resource Engineering,
The University of Western Australia, 35 Stirling Highway, Crawley
WA 6009. Email: elkington@civil.uwa.edu.au
2. MAusIMM, Associate Professor, School of Civil and Resource
Engineering, The University of Western Australia, 35 Stirling
Highway, Crawley WA 6009. Email: durham@mining.uwa.edu.au
3. MAusIMM, Group General Manager – Engineering, Snowden
Mining Industry Consultants, 87 Colin Street, West Perth WA 6005.
Email: pmyers@snowdengroup.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
creating a circular problem. Lane suggested that a declining
cut-off grade strategy often delivers a high NPV. Lane (1964;
1988) also found that mining costs should be incorporated in the
cut-off decision for underground mines, but not open pit mines.
As such, the cut-off grade for underground mines is generally
higher.
Nilsson (1982) proposed that optimising cut-off grade is more
important for underground mines than open pit mines, because
material left unmined underground may not be recoverable at a
later date. Additionally, underground mines will often have a
higher proportion of fixed costs. Since the allocation of fixed
costs is a contentious issue, the correct cut-off calculation for
underground mines is complex. Border (1991) recognised that
mining cost often varies spatially in underground mines and
should be reflected in a spatially varied cut-off grade. Hall and
de Vries (2003) suggested that cut-off grade could vary by
location for underground mines, even when all of the parameters
are held constant. Horsley (2005) demonstrated that low-grade
lodes can often be extracted profitably at a lower cut-off grade if
there is a nearby higher grade lode that can carry the fixed and
development costs. Additionally, varying cut-off grade by
location can help to improve equipment utilisation and smooth
production schedules over the long-term.
Heuristic procedures for the automation of stope design were
proposed by Alford (1995) and Thomas and Earl (1999). These
approaches are based on ‘floating’ a minimum sized mining
shape over the resource to determine, at a given cut-off grade,
which material should be included in the optimal mining outline.
These tools do not provide a stope design, but can be used as a
guide to indicate the portion of the orebody, inclusive of planned
dilution, to include in the stope design. Ovanic and Young (1995)
used separable programming to optimise the placement of stopes
within a row of stope rings. Grieco (2003, 2007) used a mixed
integer programming (MIP) formulation, incorporating grade
uncertainty, to optimise stope design for a given risk profile.
Brazil et al (2007) described a tool for optimising the overall
development network. This approach included the incorporation
of Steiner vertices to the network to minimise a combined
haulage and development cost objective function, given a series
of access points with associated tonnages. The authors also
developed a tool for optimising the path of the decline through
the optimised network with a dynamic programming approach
based on the optimised network. Kuchta, Newman and
Topal (2003) used a MIP formulation to improve medium-term
scheduling at the Kiruna sublevel caving operation. The schedule
was production based, with the objective being a combination of
minimising costs and deviations from desired targets. This was
done by optimising the timing and placement of LHDs within the
length of the mine. A MIP formulation for sublevel stoping was
proposed by Nehring and Topal (2007). The authors incorporated
activities other than production in the optimisation, such as stope
preparation, void and filling tasks, as well as constraining the
mine against multiple fill exposures.
Cut-off grade optimisation and schedule optimisation for a large
sublevel open stoping mine was undertaken by Smith and
O’Rourke (2005). They applied a mixed integer (MIP) formulation
that simultaneously designed and scheduled stopes for a large
orebody. The resulting solution was found to have a NPV 29.5 per
cent higher than that found using conventional methods, a
substantial improvement. They also found that the life-of-mine
optimisation was able to achieve the target production rate for
longer than if planned with a rules-based scheduler.
Spectrum Series Volume 17 241
T ELKINGTON, R DURHAM and P MYERS
Typically, underground strategic mine plans are completed
with the sequential determination of each aspect in isolation.
Because this approach does not accurately model many
important relationships between aspects, the resultant mine plan
might be a feasible solution, but is unlikely to find the theoretical
‘optimal’ solution. Without a method to find the optimal solution,
the level of suboptimality cannot be quantified. Additionally, the
time- consuming and manually intensive nature of this process
reduces the ability to complete multiple iterations efficiently.
Currently, there is no generalised tool or algorithm available to
optimise each aspect of the mine configuration simultaneously.
Optimising multiple aspects simultaneously results in an increase
in the number of possible solutions. Carter, Lee and Baarsma
(2004) suggested:
the inherent complexity of the underground
problem, will in all likelihood defy the
development of a single solution method
analogous to the open pit approach.
While this is the ultimate goal for strategic underground mine
planning, this paper focuses on an approach to simultaneously
incorporate information from multiple aspects of the mine plan
into a MIP formulation. The resultant optimisation may not find
the theoretical ‘optimal’ solution, but supports informed decision
making. A hypothetical case study is used to demonstrate the
application of the approach.
MIXED INTEGER PROGRAMMING
Mixed Integer Programming (MIP) is a part of the Linear
Programming (LP) family. This mathematical optimisation
technique solves for a number of decision variables. Each of
these variables is given a value and is incorporated within
linearly expressed constraints. The most appealing property of
this technique is its ability to find the optimal solution and as a
result, solutions are able to be compared consistently.
MIP formulations have been used widely in open pit
optimisation. Ramazan and Dimitrakopoulos (2007) presented
the generalised MIP formulation for optimising NPV for open
pits and tested various adjustments with respect to performance.
Caccetta, Kelsey and Giannini (1998) presented a Lagrangian
relaxation of the MIP formulation in which mining and
processing constraints are moved into objective function as ‘soft’
constraints. Caccetta and Hill (2003) outlined a branch and cut
procedure for providing more efficient convergence of a similar
formulation.
Significant work has also been completed in regards to
optimising open pits for risk, in addition to value.
Dimitrakopoulos and Ramazan (2004) outlined a formulation of
a MIP approach which included grade uncertainty. In this
method, the objective was to minimise a cost function comprised
of applying penalty costs to the probability that a block will have
a grade less than required. A stochastic integer programming
(SIP) formulation was proposed by Ramazan and
Dimitrakopoulos (2007). This method accounted for the risk of
violating tonnes and grade constraints when considering multiple
orebody realisations. Menabde et al (2007) provided a MIP
formulation for a number of orebody realisations and showed
significant average value increases over traditional deterministic
approaches.
METHOD
The method proposed does not promise to provide the ‘optimal’
mine plan for a given project configuration. There are too many
permutations to consider. The method used for this paper
considers how information from all aspects of the strategic mine
plan for a given configuration can be combined into a
mathematical MIP model. The solving of this mathematical
242 Spectrum Series Volume 17
model can provide important information to support strategic
decision making prior to proceeding to detailed design.
The basis of this approach is that many of the aspects of the
strategic mine plan are dependent on each other. For example,
stope design is dependent on the resource geometry and ore
distribution, the mining method, the geotechnical constraints and
the cut-off grade. Additionally, development design is dependent
on the stope design and development configuration and
geotechnical constraints. This method is based upon an assumed
deterministic block model and set of economic and technical
parameters. Future work to identify methods to incorporate
additional information regarding uncertainty would likely lead to
a more robust strategy under a range of conditions.
Stope design related to cut-off grade
The first step in preparing the MIP model is to determine the
stoping outlines for a given mining configuration and set of
geotechnical constraints at a range of cut-off values (usually
grade). Snowden’s Stopesizor (Myers et al, 2007) software is
used for this purpose. The heuristic stope optimisation algorithm
involves floating a Selective Mining Block (SMB) over the
orebody. The algorithm searches for the SMB that achieves the
maximum average value for the given resource. This SMB is
then ‘flagged’ to be mined. The algorithm then searches for the
SMB that achieves the next highest average value (omitting
already flagged blocks). This process continues until there are no
more SMBs that can be added above a given minimum cut-off
value. Flagged blocks are allocated to cut-off ‘bins’ according to
the order they are included in the mining outline. A consideration
of the algorithm is the concept of ‘partial’ SMBs. Partial SMBs
share blocks with already flagged SMBs. The algorithm does not
include the grade of previously flagged blocks in the average
grade calculation for the partial. Hence, all blocks entering the
mining outline must incrementally meet the cut-off value. The
algorithm outputs a modified block model coded with the cut-off
grade at which the block is included in the mining outline
(Figure 1).
A
B
FIG 1 - Plan section of an example (A) resource model and
(B) mining outline (coloured by cut-off grade).
An appealing property of the algorithm (exploited in the
approach for this paper) is that mining outlines are nested
according to cut-off grade, ie high cut-off outlines are nested
within low cut-off outlines. This result is analogous to the
parametric generation of shells used for open pits (Whittle,
1991). As this method does not account for geotechnical
Advances in Orebody Modelling and Strategic Mine Planning I
 OPTIMISING VALUE FOR AN UNDERGROUND PROJECT CONFIGURATION
maximum stope size constraints or support pillars. As a result,
the algorithm does not produce actual stopes; rather outlines on
which to base practical stope designs. The algorithm does not
incorporate economic and technical parameters and hence does
not give the optimum mining outline, only a series of outlines to
be considered for further evaluation based on specified cut-off
values. However, a grade-tonnage curve based upon the stope
optimisation results provides a more accurate set of data for
economic evaluation than does the raw resource model
(Figure 2). This is largely due to the inclusion of planned dilution
and the exclusion of isolated above cut-off value resource blocks.
FIG 2 - Resource – mining inventory comparison (annotated by
cut-off grade).
Development design related to cut-off grade
Development design is strongly dependent on stope design, or
mining outline in this case. Conceptual development designs
were created for each stope design (one stope design for each
cut-off grade). These designs were completed manually, based on
a given development configuration. An extension to this study
would be to optimise the development network/decline using
methods such as those outlined by Brazil et al (2007). The
development designs were completed in order to determine the
relationship between development metres and cut-off grade on
each level (Figure 3). This allows a given stope design to be
matched with a valid development design.
Strategic schedule optimisation
The schedule optimisation used in this method incorporates the
information gained through stope and development design
analysis at a number of cut-off grades. It is assumed that for a
given project configuration, development design is dependent on
stope design and stope design is dependent on cut-off grade.
FIG 4 - Optimisation concept.
Advances in Orebody Modelling and Strategic Mine Planning I
A
FIG 3 - Level development design for (A) 1 g/t cut-off and
(B) 7 g/t cut-off.
Hence, the global optimisation focuses on cut-off grade and
schedule optimisation, whilst ensuring that the stope and
development designs are consistent with the cut-off grade(s)
selected in the optimisation (Figure 4). The objective of the
optimisation is to select the cut-off grade policy and schedule
that maximises NPV, ie what should be mined and when it
should be mined. It is recognised that NPV might not be the
primary objective of every underground operation, and that
optimisations based on other objectives such as total cash flow,
mine life, or risk could be considered.
Completing a multi-period MIP optimisation on a block-
by-block resolution is beyond current hardware and software
capabilities. In order to make it practical to solve the problem, it
is necessary to create a smaller number of activities to schedule.
This requires the grouping of blocks. An example of this is given
in Stone et al (2007) for an open pit application. Grouped blocks
will carry the same (averaged) properties, and resolution is lost.
Grouped blocks also report the same output, including their
ultimate destination and the time period in which this occurs.
Because of this, it is important that resolution is managed
efficiently by grouping blocks that are likely to be mined
Spectrum Series Volume 17 243
T ELKINGTON, R DURHAM and P MYERS
together and sent to the same destination. The grouping process
used for the case study in this paper is to group blocks by level
and cut-off grade to form stoping activities. It is assumed that
material on a given level will be extracted at approximately the
same time, and that material in each cut-off grade interval will be
sent to the same destination. Grouping is also applied for
development activities. Decline development and level access
development (access crosscut and footwall drive) is grouped by
level. Finally, variable development (development driven by
tonnage/cut-off grade) is incorporated within the stoping
activities.
For the MIP model, a decision variable is created for each
stoping and development group in each period. Each decision
variable measures the proportion of the activity that is completed
in the given period. NPV theory is based upon the concept that
cash flow received today is more valuable than an equivalent
cash flow received in the future. As the objective of the MIP
optimisation is to maximise NPV, the objective value associated
with each activity is equal to the cash flow (positive or negative)
discounted at a specified rate. The cash flow associated with each
activity is calculated from the economic and technical
parameters associated with the activity. A range of important
constraints are applied in order to reflect the mining practicality
of the given project configuration, as well as to control the
optimisation.
The constraints that reflect the project configuration include a
minimum and maximum limit on mining/processing tonnages and
development metres per time period. Note that for a configuration
with a single processing plant, the mining and processing limits
would usually be the same or similar. Development constraints can
be grouped according to vertical and horizontal development to
reflect the fact that different equipment is sometimes used for each
purpose. These constraints can be applied globally or locally (to a
subset of activities). For example, while the overall mining limit
might be 1 Mt/a, the maximum that can be extracted from a single
mining area might only be 500 kt/a. Additionally, blending
constraints and selling constraints can be applied to enforce head
grade bounds and total metal output. Wooller (2001) provides an
interesting implementation of selling constraints to model
processing plant performance.
The project configuration will also outline the direction of
mining. This is imposed as a series of sequencing constraints
based on precedence relationships between activities. These
relationships can link development and stoping activities,
development activities only or stoping activities only. While
geotechnical constraints are not directly applied in this method,
sequencing constraints can be used to simulate this. Integer
decision variables are introduced to monitor whether all activities
precedent to a given activity have been completed.
The optimisation is also controlled by two other important
types of constraint. Firstly, constraints are imposed to ensure
each activity is only completed once. Secondly, constraints are
used to allow only one cut-off grade to be selected for a given
activity. Integer decision variables are used to manage these
constraints – the solution to the optimisation problem can be
interpreted into a mining and development schedule.
Consequently, each block can be coded with the time period in
which it is mined. This enables the output to be exported to mine
design or scheduling packages for further detailed analysis.
CASE STUDY
The case study project used to demonstrate the application of the
approach is a hypothetical gold project.
Background
The orebody is a large dual lode gold deposit (Figure 5a). The two
lodes are near vertical and cover 900 m in vertical extent. The
orebody lies beneath 60 m of overburden. The lodes average 40 m
244 Spectrum Series Volume 17
A B
FIG 5 - View of orebody looking south coloured by (A) grade,
(B) mining front.
in thickness with a minimum separation of 40 m between the
lodes. The rock conditions, vertical extent and close proximity of
the lodes suggest that open stoping would be an appropriate
mining method. Mining inventories incorporating planned dilution
were generated for each cut-off grade with Snowden’s Stopesizor
software tool. Cut-off grades considered in this analysis were 1 g/t
to 10 g/t in 0.5 g/t increments. The minimum mining shape was
assumed to be 20 mE × 20 mN × 20 mRL, based on preliminary
geotechnical analysis. Applying this shape to the orebody
resulted in significant planned dilution and the divergence of the
mining inventory from the resource (Figure 2). This result
reflects the bulk nature of the mining method.
A basic development design was completed on each level for
each cut-off grade (Figure 3). The decline was placed on the
western side of the orebody and the ore was accessed via a drive
located centrally to the two lodes. The designs were used to
derive average decline metres and access development for each
level, as well as the relationship between the tonnes in the
mining inventory and the development required on each level.
Decline and level access development was assumed to be
independent of cut-off grade.
The underground optimisation was based upon optimising
value for a 1 Mt/a production rate. The overall mine extraction
sequence is top-down via a bottom-up stoping method from four
production fronts, each capable of providing 500 ktpa.
Consequently, at least two fronts need to be operational in any
period to achieve full production. The top level in each
production front remains as a pillar, which can be extracted once
all levels above it are completed. A maximum decline
development rate of 1500 M/a and horizontal development rate
of 5500 M/a was assumed. Development was constrained so that
the decline and level access must precede any level development
on a given level.
An initial optimisation was conducted without any assumed
production fronts or stoping activity sequencing. This allowed
the optimisation to reveal the areas where the highest values
exist. These areas indicate the best locations to start production
from, if geotechnical constraints were not important. This
represents an upper bound solution to guide future strategy.
Production fronts were allocated based on the concept of trying
to replicate the upper bound solution with practical mining
Advances in Orebody Modelling and Strategic Mine Planning I
 OPTIMISING VALUE FOR AN UNDERGROUND PROJECT CONFIGURATION

constraints. This was done by starting production for a

production front on a level that is extracted early in the upper
bound solution. Likewise, crown pillars for each production front
were located at levels taken late in the best case solution.
Production front allocation is shown in Figure 5b.

Results
This project configuration was analysed under two different
assumptions regarding cut-off grade policy. Firstly, the
configuration was optimised for each cut-off grade, assuming that
the same cut-off grade is applied on every level. Secondly, the
configuration was optimised whilst allowing cut-off grade to be
varied between levels. An important assumption made for this
optimisation was that the stope design on a level is independent of
the design of all others. This allowed for the case of varying
cut-off grade by level to be analysed. Additionally, the
conservative assumption was made that once mining is completed
FIG 7 - Single cut-off grade 5 g/t production schedule.
on a level to the selected cut-off grade, the material beneath this
cut-off grade is sterilised.

Single cut-off grade

A single cut-off is perceived to be the simplest approach in terms
of design and operation. Figure 6 shows that NPV is maximised
at a cut-off grade of 5 g/t with 19.3 Mt of ore.

FIG 8 - Cumulative discounted cash flow schedule comparison.

significant. Though this will result in a shorter mine life, the

FIG 6 - NPV curve.
The break-even cut-off, most commonly applied in operating
practice, is 3.5 g/t. At the break-even cut-off grade there is a
17 per cent drop in NPV accompanying a 4.6 Mt increase in ore
processed. In this case, the break-even cut-off strategy promotes
mining longer for less NPV. The schedule, at a 5 g/t cut-off, is
shown in Figure 7. One of the key results of the schedule
optimisation is that production front four does not commence
until year 11 of the operation. The decline is capable of reaching
this front by year six. The optimisation was able to resolve the
compromise between grade differentials and the added cost of
development in deciding to delay this development. An
additional characteristic of the optimisation was the ability to
maintain a sustainable production rate until the end of the mine
life. This was due to the multi-period optimisation being able to
identify and rectify bottlenecks by adapting the strategy in early
periods.
The source of the increased value associated with an increase
in cut-off grade above break-even is opportunity cost.
Discounting makes cash inflows in later periods (of a long life
operation) worth little in today’s terms. Figure 8 shows the
minimal value contribution of the final five years of the
break-even option. Hence, it is often (and demonstrated in this
case) appropriate to raise the cut-off grade to achieve a higher
average grade in those periods when discounting is not as
reduction is taken from periods that do not provide meaningful
contributions to discounted value. It is expected that the shorter
the mine life, the closer the optimal single cut-off grade will be
to the break-even cut-off grade.
Spatially varying cut-off grade
Testing Horsley’s (2005) findings that value can be improved by
varying cut-off grade spatially, the optimisation was allowed to
allocate any of the 19 cut-off options on each level – this
optimisation resulted in an NPV improvement of seven per cent
(Figure 6) over the single 5 g/t cut-off grade. The tonnage
associated with this option is 17.4 Mt. The seven per cent
increase in value is gained through mining at a higher average
grade in periods 4 - 11 (Figure 9). Periods 1 - 3 are constrained
by the development to areas that can be mined and hence average
grade is low. The higher average grade, due to increased cut-offs
in the periods 4 - 11, results in greater cash flows when
discounting does not impact so greatly. At the end of the mine
life, additional material is able to be mined at a lower cut-off
grade than 5g/t in order to maximise extraction, due to the lesser
impact of discounting.
Overall, this solution represents a 24 per cent increase in value
over the break-even solution with a mine life of six years less. This
is assuming that the break-even schedule has been optimised for
NPV, which it rarely is. This method could potentially find a
greater NPV improvement over a non-optimised break-even plan.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 245
T ELKINGTON, R DURHAM and P MYERS
FIG 9 - Spatially varying cut-off grade production schedule.
Lane (1964; 1988) suggested that cut-off was time-dependent
rather than spatially dependent. Figure 10 shows that levels that
were mined early had a much higher cut-off grade (up to 7.5 g/t)
than those level mined later (down to 3.5 g/t), due to the effect of
discounting. The only exception is the first few years, which
have lower cut-offs due to the lack of available ore. The general
trend for each production front was for the cut-off grade to
decrease as mining proceeds. In fact, plotting the start time
against cut-off grade for each level resulted in a clear negative
linear relationship (Figure 11). This indicates that in this case,
time is a more important driver of cut-off grades than location.
Production front one has much lower cut-off grades due to the
FIG 10 - Spatially varying cut-off grade – optimal solution.
FIG 11 - Relationship between start time and cut-off grade (one
sample from each level).
246 Spectrum Series Volume 17
operation being development-constrained at the time this area is
to be mined.
MULTIPLE CONFIGURATION ANALYSIS
The concept of rapid optimisation of value for a single project
configuration allows a number of configurations to be
subsequently optimised and compared in an efficient manner. The
following analyses describe how the optimisation methodology
can be extended to allow for greater information in strategic
decision making.
Production rate optimisation
Production rate has a number of impacts for a new underground
project. An increase in production rate can result in an increase
in upfront capital cost due to plant and mining fleet size, a
decrease in unit mining and processing costs due to economies of
scale, an increase in fixed costs associated with maintaining
larger fleet and plant, an increased presence of bottlenecks due to
additional equipment, an increased effort to maintain a
sustainable production rate and a shortened mine life. Figure 12
shows a contour plot interpolating NPV for each combination of
19 cut-off grades (single cut-off) and eight production rates
between 0.8 Mt/a and 1.5 Mt/a. The optimal NPV is achieved at
a cut-off grade of 5.5 g/t and production rate of 1.3 Mtpa
(marked with a ‘×’ in Figure 12). The contour line immediately
around that point represents the solutions within five per cent of
the optimal solution and indicates a number of acceptable
solutions in this case. The break-even cut-off strategy of 3.5 g/t
does not lie within this optimal region. In fact, depending on the
production rate, this strategy will be somewhere between ten per
cent and 35 per cent suboptimal.
FIG 12 - Interpolated relative NPV.
Strategy sensitivity to price
The most relevant question in terms of price sensitivity is:
which strategy, selected today, allows the project
to be optimal or near optimal under a range of
commodity price scenarios?
The ability to quickly optimise for a single project configuration
enables a number of commodity price scenarios to be considered.
Scenarios of price between -30 per cent and +30 per cent (relative
to the base case price) were considered for each of the 19 cut-off
grade options (single cut-off). The results (Figure 13) are plotted
in a contour graph, with an interpolation of NPV relative to the
maximum NPV for each price scenario. An important result of this
analysis is that the selection of a cut-off grade of 5.5 - 6 g/t would
be, at most, five per cent suboptimal under all of these price
conditions. This cut-off grade would be considered quite resilient
to price uncertainty. The modest negative relationship between
Advances in Orebody Modelling and Strategic Mine Planning I
 OPTIMISING VALUE FOR AN UNDERGROUND PROJECT CONFIGURATION
FIG 13 - Interpolated NPV (relative to the maximum NPV at
each price).
price and optimal cut-off grade indicates that for this case, an
optimised mine strategy is quite resilient to changes in price. The
break-even cut-off grade strategy of 3.5 g/t for the base case is
between 11 per cent and 100 per cent suboptimal under the
various price conditions. A break-even strategy not only misses
value opportunity, but is also a higher risk option.
CONCLUSIONS
This paper presents an approach to incorporating information from
a number of aspects of underground strategic mine planning into a
single optimisation. The case study found that substantial gains in
value could be accessed by mining in a more strategically focused
way. In particular, it showed that a break-even cut-off grade
strategy encourages mining more material for less value. Gains in
value of up to 24 per cent were found through optimising cut-off
grade spatially and with time. This maximum gain was achieved
while mining for six years less. Comparatively greater gains were
likely, given that traditional break-even cut-off grade schedules are
not optimised.
The optimal mine plan for the base case was also found to be
much more resilient to price uncertainty than a break-even cut-off
strategy. Under a ±30 per cent price range, the optimal base case
cut-off is at most five per cent suboptimal. For the same range, the
break-even strategy is between ten per cent and 100 per cent
suboptimal. The gain in value and reduction in risk were made
with little, if any, additional planning effort. This gain does not
include the potential for additional improvements to be made from
considering other configurations that could increase value or
mitigate risk. Given the current economic climate, the need to
optimise the utilisation of mineral deposits has become essential.
Optimisation methods such as the one described in this paper
should be applied to all mining operations. They allow many
project configurations to be compared on a consistent basis, and
provide confidence that the best strategy has been found. The
optimal solution can then be used to drive more detailed design
and scheduling.
ACKNOWLEDGEMENTS
The authors would like to thank Snowden Mining Industry
Consultants for their support of this research and for applying a
similar approach for their in-house Evaluator tool. Additionally,
we would like to thank R Secis and A Earl from Snowden
Mining Industry Consultants for reviewing this paper.
REFERENCES
Alford, C, 1995. Optimisation in underground mine design, in
Proceedings APCOM XXV, Brisbane, pp 213-218.
Berry, E, 1922. Present value in its relation to ore reserves, plant capacity
and grade of ore, American Institute of Mining and Metallurgical
Engineers, July, 11-16.
Advances in Orebody Modelling and Strategic Mine Planning I
Border, S, 1991. Optimisation of cut-off grades during design of
underground mines, in Proceedings Mining Industry Optimisation
Conference, pp 21-27 (The Australasian Institute of Mining and
Metallurgy: Melbourne).
Brazil, M, Lee, D, Rubinstein, J, Thomas, D, Weng, J and Wormald, N,
2007. Optimisation in the design of underground mine access, in
Orebody Modelling and Strategic Mine Planning, second edition (ed:
R Dimitrakopoulos), pp 141-144 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Caccetta, L and Hill, S, 2003. An application of branch and cut to open
pit mine scheduling, Journal of Global Optimisation, 27:349-365.
Caccetta, L, Kelsey, P and Giannini, L, 1998. Open pit mine production
scheduling, in Proceedings Third Regional APCOM Symposium –
Computer Applications in the Minerals Industries International
Symposium, pp 65-72 (The Australasian Institute of Mining and
Metallurgy: Melbourne).
Carter, P, Lee, D and Baarsma, H, 2004. Optimisation methods for the
selection of an underground mining method, in Proceedings Orebody
Modelling and Strategic Mine Planning, pp 7-12 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Dimitrakopoulos, R and Ramazan, S, 2004. Uncertainty-based production
scheduling in open pit mining, Transactions of the SME, pp 106-112.
Grieco, N, 2003. Risk analysis of optimal stope design: Incorporating
uncertainty, Master of Philosophy thesis, The University of
Queensland.
Grieco, N and Dimitrakopoulos, R, 2007. Grade uncertainty in stope design:
Improving the optimisation process, in Orebody Modelling and Strategic
Mine Planning, second edition (ed: R Dimitrakopoulos), pp 167-174
(The Australasian Institute of Mining and Metallurgy: Melbourne).
Hall, B and De Vries, J, 2003. Quantifying the economic risk of
suboptimal mine plans and strategies, in Proceedings Mining Risk
Management Conference, pp 59-69 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Horsley, T, 2005. Differential cut-off grades, in Proceedings Ninth
Underground Operators’ Conference, pp 103-109 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Kuchta, M, Newman, A and Topal, E, 2003. Implementing a production
schedule at lkab’s Kiruna mine, INTERFACE, 34(2)124-134.
Lane, K, 1964. Choosing the optimum cutoff grade, Quarterly for the
Colorado School of Mines, 59(4):811-829.
Lane, K, 1988. The Economic Definition of Ore: Cut-Off Grades in
Theory and Practice (Mining Journal Books: London).
Menabde, M, Froyland, G, Stone, P and Yeates, G A, 2007. Mining
schedule optimisation for conditionally simulated orebodies, in
Orebody Modelling and Strategic Mine Planning, second edition (ed:
R Dimitrakopoulos), pp 379-384 (The Australasian Institute of Mining
and Metallurgy: Melbourne).
Myers, P, Standing, C, Collier, P and Noppe, M, 2007. Assessing
underground mining potential at Ernest Henry mine using
conditional simulation and stope optimisation, in Orebody
Modelling and Strategic Mine Planning, second edition (ed: R
Dimitrakopoulos), pp 191-200 (The Australasian Institute of Mining
and Metallurgy: Melbourne).
Nehring, M and Topal, E, 2007. Production schedule optimisation in
underground hard rock mining using mixed integer programming, in
Proceedings Project Evaluation 2007, pp 169-178 (The Australasian
Institute of Mining and Metallurgy: Melbourne).
Nilsson, D, 1982. Optimum cut-off grades in underground mining,
Canadian Mining Journal, 103(10):65-70.
Ovanic, J and Young, D, 1995. Economic optimisation of stope geometry
using separable programming with special branch and bound
techniques, in Proceedings Third Canadian APCOM, Montreal,
pp 22-25.
Ramazan, S and Dimitrakopoulos, R, 2004. Recent applications of
operations research in open pit mining, Transactions of the SME,
314:73-78.
Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
long-term production scheduling for open pit mines with a new integer
programming formulation, in Orebody Modelling and Strategic Mine
Planning, second edition (ed: R Dimitrakopoulos), pp 385-392 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Spectrum Series Volume 17 247
T ELKINGTON, R DURHAM and P MYERS

Smith, M and O’Rourke, A, 2005. The connection between production Whittle, J and Rozman, L, 1991. Open pit design in the 90s, in
schedules and cut-off optimisation in underground mines, in Proceedings Mining Industry Optimisation Conference, pp 13-19
Proceedings APCOM XXXII. (The Australasian Institute of Mining and Metallurgy: Melbourne).
Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse, Wooller, R, 2001. Cut-off grades beyond the mine – Optimising mill
P, 2007. Blasor – Blended iron ore mine planning optimisation at throughput, in Mineral Resource and Ore Reserve Estimation – The
Yandi, in Orebody Modelling and Strategic Mine Planning, second AusIMM Guide to Good Practice (ed: A Edwards), pp 459-468 (The
edition (ed: R Dimitrakopoulos), pp 133-136 (The Australasian Australasian Institute of Mining and Metallurgy: Melbourne).
Institute of Mining and Metallurgy: Melbourne).
Thomas, G and Earl, A, 1999. The application of second-generation
optimisation tools in underground cut-off grade analysis, in
Proceedings Strategic Mine Planning, pp 175-180 (Whittle
Programming Pty Ltd: Box Hill).

248 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I

Classification of Mining Methods for Deep Orebodies
V Oparin1, A Tapsiev2 and A Freidin1
ABSTRACT
Classifications of mining methods date back to the 1960s and use the
following criteria: type and size of a mineral deposit, mined-out space
support, state of a working excavation, type of a face, roof support, etc.
Since that time mobile mining machinery has appreciably advanced, and
some mining methods have lost their importance. Mobile mining
equipment is effective in definite mining methods: in open mined-out
space, mining with backfill and combined schemes. The transition to
deep mining inevitably results in the sharply worsened technical and
geomechanical conditions, and some well-established techniques for safe
mining at rock burst hazardous deposits can not be considered
self-sufficient. This paper puts forward a classification of deep mining
methods, considering rock pressure control. The classification involves
three classes: mining with backfill, mining with caving of overlying rocks
and combined mining with backfill and caving.
INTRODUCTION
Many foremost Russian and foreign researchers and mining
engineers have been scrutinising methods of grouping existing
underground mining methods. More than 60 published
classifications have undergone critical analysis in fundamental
monographs by Trushkov (1947), Agoshkov (1965), Baikonurov
(1969) and Imenitov (1978). Trushkov conceded on inexpediency
to try and embrace all coal and ore mining methods in a uniform
classification due to ‘crockness and inconsistency of these
methods, and their not-hands-down use in practice’ (Trushkov,
1947). Unlike coal formations, orebodies have multiple sizes and
shapes, occurrence conditions, physico-mechanical properties,
mineral composition and mineral values, which gave rise to a
range of 170 to 200 variants of ore mining schemes applied in
underground mines in the 20th century (Trushkov, 1947;
Agoshkov, 1965).
The totality of underground ore mining methods caused
difficulties with scientific grouping, however, it also fostered
many characteristics for classifying them. Different researchers
assumed different classification criteria (please refer to
Baikonurov, 1969); for example, the type and size of a deposit
(Crane, G J Young, I Pokrovsky, G E Bakanov and L I Baron);
mined-out space support (F W Sperr, Y H Rayt, E C Mitke and V
N Semevsky); working excavation support (J F Clelland, C F
Sackson and E D Gardner, USA Mining Bureau’s classification,
N I Trushkov, N A Starikov, G N Popov and V R Imenitov);
working excavation state (M I Agoshkov and R P Kaplunov);
mined-out space support during mining and backfill (V T
Markelov); type of a face, roof support, block undercutting
technique (American Institute of Mining Engineers and
Metallurgists); direction of mining advance and backfill method
(Baikonurov, 1969); mining stages (Steshenko). The above-listed
criteria rank and characterise mining methods in varying degrees
and yet are of little use in actual comparison and selection. In
Russia, the classifications by Agoshkov (1965) and Imenitov
(1978) have found deserved recognition and wide application in
underground mine planning. Agoshkov grouped mining methods
by ‘the state of the working excavation in the course of actual
mining’ and Imenitov used the criterion of ‘working excavation
1. Institute of Mining, Siberian Branch, Russian Academy of Sciences,
54 Krasny Prospect, Novosibirsk 630091, Russia.
2. Institute of Mining, Siberian Branch, Russian Academy of Sciences,
54 Krasny Prospect, Novosibirsk 630091, Russia.
Email: atapsiev@misd.nsc.ru
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
support method in the course of mining’. We acknowledge the
effectiveness and advantages of both classifications, though some
comments should be made. First, these classifications appeared
in the mid-1960s. Mining methods have been continuously
evolving since then, and many of them have dwindled, especially
for medium thick and very thick ore deposits. These days,
underground mining practice involves lively mobile machinery
sets, which has drastically simplified preparatory and
development operations as well as stoped excavation, boosted
mining works and minimised hand-labour amount due to the
overall mechanisation.
Highly laborious mining with portable equipment is now used
in the excavation of very thin orebodies with high useful mineral
content. This method and equipment may only be applied for
actual mining of medium thick and very thick ore deposits if
low-paid manpower is in employment, mineral resources are of
low value and the total enterprise is of no consequence to be
re-constructed or technically modernised. The analysis of the
mobile machinery operation shows that they are efficient in a
limited number of mining methods, namely:
• mining with backfill (longwall and chamber mining,
including the chamber-and-pillar method);
• mining without backfill (chamber-and-pillar and flat-bottom-
chamber mining);
• mining with ore and rock caving (block caving and,
especially, sublevel caving with frontal and frontal-areal ore
drawing); and
• combined mining (with backfill and caving).
The gained engineering level of underground mining with
mobile machinery and successful robotic automation of ore
excavation processes at some enterprises (Bronnikov, 1982;
Oparin, 2007) will lead to further universalisation of mining
methods. It is also important to consider that underground mining
gets deeper by 20 - 40 m annually, and now reaches nearly a four
km depth in South Africa and India. More than ten underground
mines in USA, Canada and South Africa are below 2.5 km depth.
Deep mining conditions are very difficult: rock temperature is
higher, the physico-mechanical properties of rocks change, angles
of rock mass movement flatten and rock pressure grows and
manifests itself dynamically. More and more deposits acquire the
category of rock burst hazardous. Mining theory and practice have
elaborated measures for minimising the hazard of rock pressure
(Bronnikov, 1982) to:
• prepare and develop orebodies in a minimum fractured rock
mass so that no stress concentration areas arise (pillars,
overhangs) resulting in burst-like rock failure;
• orient development workings and stopes along the maximal
effective stresses; and
• to advance stoped excavation as divaricated fronts, or as a
single front towards the wing of a level.
These measures collide with many methods of underground
mining, for example, with ‘ore shrinkage’ and ‘mining without
backfill’ by Agoshkov (1965), or ‘mining with naturally
supported working excavation’ by Imenitov (1978).
Different methods of mining with ore shrinkage that are in wide
application in thin and steep orebodies, involve ascending stoping
towards a diminished pillar. This is not permissible for us in rock
burst hazardous rock masses. For instance, at Yuzhny complex
deposit (GMK Dalpolimetal Co), where thin and strong ore veins
Spectrum Series Volume 17 249
V OPARIN, A TAPSIEV and A FREIDIN
occur in hard host rocks prone to elastic energy accumulation and
dynamic failure (Pilenkov, 1990), series of rock bursts with severe
after-effects took place at a depth of 150 - 200 m from the surface
when mining with ore shrinkage approached the crown pillar. The
ground control is extremely limited when mining is carried out
without backfill, and the roof is supported with ore pillars and
lining. In this case, the growing mined-out space loses its
equilibrium, dynamic rock failure conditions originate and
large-scale rock fall becomes highly possible. A representative
illustration of the above is the underground mines at Zhezkazgan
cupriferous sandstone deposit (Yun, 1997), where the chamber-
and-pillar method involves retaining enlarged inner-chamber and
panel pillars (up to 40 per cent and above of a panel area).
Nonetheless, starting from a 450 m depth, falls and burst-like
failure of pillars occur again and again. The calculations indicate
that for chamber-and-pillar mining at a depth of 1000 m, more
than half-reserves should be left as pillars for roof support safety
(Bronnikov, 1982; Zamesov, 1979).
METHOD OF CLASSIFICATION
An important point in Agoshkov (1965) and Imenitov’s (1978)
mining method classifications is the lack of the systems of
chamber mining with backfill. These systems are of use in
longwalling and chamber- and-pillar working at 1000 m depth and
deeper. The underground chamber-and-pillar mining with backfill
is applied extensively in Canada (Inco, Geco, Kidd Creek Mines),
Finland (Outokumpu) and Australia (Mount Isa) (Khomyakov,
1984), as well as in the former USSR countries (Gaisky,
Leninogorsky, Zyryanovsky Mines, etc). Some engineers think
these geo-technologies are useful in mining no deeper than 600 -
800 m, due to the growing bearing pressure after the first
chambers are extracted (Bronnikov, 1982; Zamesov, 1979). Failure
of pillars and excavations at ore drawing-off level necessitates
introduction of longwall chamber methods that proved themselves
in the Norilsk mines. Thus, for rock burst hazardous ore deposits,
we think it is possible to characterise the allowable mining
systems by:
• the principal ground control method for the mined-out space,
• the stoping excavation advance direction, and
• the ore breaking and drawing method.
The ground control is the governing factor for any deep mining
technology. It is always present in the description and in the name
of a mining method. We assume it the basic classification
criterion.
It is natural that the principal ground control method is
contributed to by the known design solutions oriented to enhance
strain capacity of a rock mass and to monitor stresses in the
near-face area. Such supportive measures include, for example,
overworking (undermining) of the orebody, advance boring of
larger diameter holes, confined explosions, directional hydro-
fracturing, etc. These measures are feasible with any mining
method and are not the attributes of a particular recovery
mechanism. The second-listed characteristic is often present in the
name of a mining method and gives a description of it. For
example, the horizontal slicing with backfill may be ascending and
descending. Steep rock burst hazardous ore veins are mined with
sublevel stoping to the dip or on the strike. The third characteristic
in our list details the way of ore drawing, namely, areal (from
inclines), frontal or frontal-areal. Pillar mining may involve VCR,
parallel or fan-patterned boreholes. So, the second and third
criteria can and must serve for the grouping of the mining
methods.
We are proposing a variant of classification of the mining
methods for rock burst hazardous ore deposits based on the
following principles:
• the mining methods should be obligated to meet the stringent
standards of operations in the conditions of rock burst hazard;
250 Spectrum Series Volume 17
• the classification is for industrial-scale mining, with only the
exclusion of thin ore deposits where portable mining
equipment is unavoidable due to space-limited environment;
• the secondary criteria should be disregarded, for example,
when strike-line or dip-line mining is governed not by the
mining method, but by the parameters of the deposit and its
mining-technical and geomechanical conditions;
• the systems with support are withdrawn from the
classification as all of the known mining methods involve the
support; and
• the basic classification criterion allows entering any new
mining method to the classification.
Thus, in accordance with the effective ground control methods,
we divide deep mining system into three classes:
1. Class I – mining with backfill,
2. Class II – mining with overlying (host) rock caving, and
3. Class III – combined mining with backfill and caving.
Classes I and II are commonly known from technical literature
and the practice of underground mining. We would only note the
undesirable large-scale blasting during mining with caving and
frontal ore drawing at rock burst hazardous deposits. Large-scale
explosions release huge energy, which initiates displacement of
large rock blocks, causes dynamic events in the rock mass and
failure of the weakened ore drawing-off level. These events are
frequent at the Tashtagol iron-ore deposit, where the mining
method includes block caving, vibration areal drawing of ore and
large-scale block breakage. The Kiruna Mine in Sweden has the
same rock mechanics and identical depth of mining, but with
sublevel caving, slice breaking and frontal ore drawing. As a
result, dynamic pressure manifestations are absent here.
Therefore, we include only slice breaking with frontal and
frontal-areal ore drawing into our classification (Freidin, 2008).
The technologies involved with the combined ground control
methods are as follows:
• paste filling costs are expensive, and
• mining of low and mid-grade ores calls for cheaper methods.
In particular, partial backfill is admissible in thick and very
thick gentle and weak-inclined deposits, with the purpose of
providing the roof support and smooth sagging, limited dis-
placements and strains of the host rocks. In this case, the filled
areas alternate with the areas of mining under the caved roof
formed as a stable dome. First, mining captures areas with
artificial pillars, and secondly, the areas with temporarily
retained ore pillars. The smooth sagging of the roof is ensured
with the help of the artificial pillars, the sides of which are
supported by caved overlying rocks. Parameters of these pillars
should limit the rate of the roof sagging to the values of mining
with full backfill. The width of the temporary ore pillars is
designed with allowance for the stable span. Depending on the
stability of ore and host rocks, and orebody thickness, mining
with backfill may include horizontal slicing with ascending (or
descending) pillar mining, and the temporary pillar mining may
involve sublevel caving (for very thick orebodies) or block
caving (for thick deposits). The offered classification of the
mining methods is presented in Table 1. The information in this
table in terms of mining methods is further depicted in a series of
related figures.
CONCLUSIONS
This classification of mining methods for deep orebodies has
been offered for the first time, and is based on the underground
mining practice in Russia, in particular, in Norilsk Region. The
authors think the classification may successfully be used for
selecting and optimising the extraction of rockburst-hazardous
ore and rock masses.
Advances in Orebody Modelling and Strategic Mine Planning I
 CLASSIFICATION OF MINING METHODS FOR DEEP OREBODIES

TABLE 1
Classification of mining methods for rock burst hazardous ore deposits (table includes references to the corresponding figures).

Class Group Mining Application conditions Advantages Disadvantages

method
1 2 3 4 5 6
I. Ground A. Horizontal 1. Horizontal Any geometry of a deposit. Medium Simple geometry of stoping Strict regulation of slicing.
control with slicing with slicing and and greater thickness. Average faces, high extraction index. In case of high rock pressure,
backfill backfill ascending stability ores, mid-stable to unstable Mobile machinery use. Sufficient overworking is desirable,
mining host rocks. Medium to high-grade stoping productivity. Possible which is highly laborious.
ore. Unlimited depth of application. separation of dirt rocks for Organizational difficulties in
backfill. stoping automation
(Figure 1).
2. Horizontal Any geometry of a deposit. Medium The same as in ascending The same as in ascending
slicing and and greater thickness. Very unstable mining. Safe mining in heavily mining. High labour intensity
descending and heavily dislocated ores and host dislocated orebodies. and cost of backfilling
mining rocks. High-grade ore. Unlimited (Figure 2).
depth of application. Open cast and
underground mining.
3. Horizontal Gently dipping medium thick and Ascending and descending Driving and backfill of extra
slicing with very thick orebodies. Mid-stable to mining of main reserves and mine workings in the bottom
pillars on soft unstable ores and host rocks. temporary pillars. Increased of temporary pillars. Stoping
basis High-grade ore. Depth of application mining front and more stopes in stages (Figure 3).
is to be determined. under development. High
extraction index.
4. Room work Mid-thick to very thick deposits. Higher quality and completeness Mining in stages. Higher
and slicing Any dipping. Mid-stable to unstable of pillar extraction. labour intensity and lower
ores. Mid-stable host rocks. Medium stoping production (Figure 4).
to high-grade ores. Depth of
application is to be determined.
5. Longwall Horizontal and gently dipping Selective mining. Elongated Arduous working conditions
mining orebodies. Very thin to 3 - 4 m thick stoping front. High extraction and high labour intensity. Use
ores. Any stability ores, mid-stable index. of portable equipment, limited
host rocks. Paste backfill, packs, use of mobile machinery
stony and artificial blocks. (Figure 5).
High-grade ores. Unlimited depth of
application.
B. Room 6. Continuous Medium thick to very thick deposits. High stoping productivity. High ore dilution with
work mining chamber Any dipping. Medium stable ores Automation of principal cement-containing backfill.
mining and host rocks. Depth of application operations in stoping. Limited mining front
is to be determined. Medium to (Figure 6).
high-grade ores.
7. Chamber- Medium thick to very thick deposits. High stoping productivity. Mining in stages. Higher
and-pillar Any dipping. Medium stable ores Automation of principal requirements to ore extraction
mining (areal, and host rocks. Medium and operations in stoping. Extracted standards. Probable failure of
frontal and high-grade ores. Areal, frontal and ore quality control. secondary chambers (pillars)
frontal-areal frontal-areal ore drawing depending in case of high pressure. High
ore drawing) on the orebody thickness. ore dilution (Figure 7).
II. Ground A. Block 1. Block caving Very thick, high dip outcropping or High stoping productivity. Probable hanging-up of host
control with caving with slicing close to surface ores. Limited Relatively low mining costs. rocks with following
overlying and frontal application in gently dipping and High automation of stoping and burst-like destruction of large
rocks caving (frontal-areal) inclined ore deposits with weak and use of mobile machinery. structural rocks. Forced
ore drawing unstable overlying rocks. Hard Sufficient extraction index with ventilation of ore drawing-off
medium stable and stable, low and frontal-areal ore drawing. workings (Figure 8).
medium grade ores. Limited depth Ventilation due to mine
of application. Unstable and medium depression.
stable host rocks. Ores are non-liable
to spontaneous ignition and caking.
B. Sublevel 2. Sublevel Thick and very thick, high dip Separation of dirt bands to Probable hanging-up of
caving caving with outcropping or close to surface backfill. High stoping overlying rocks, displacement
frontal orebodies. Blind deposits with productivity. Mining in with and burst-like failure of large
(frontal-areal) unstable overlying rocks. Limited many stoping faces, extracted ore structural blocks. Forced
ore drawing application in gently dipping and quality control. Automation of ventilation of ore drawing-off
inclined deposits. Medium hard and stoping. Sufficient extraction workings (Figures 9a and
hard ores, non-liable to spontaneous index with frontal-areal ore 9b).
ignition and caking. Low and drawing. Ventilation due to mine
medium grade ores. Depth of depression.
application is to be determined.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 251
V OPARIN, A TAPSIEV and A FREIDIN

TABLE 1 cont ...

Class Group Mining method Application conditions Advantages Disadvantages

3. Strike-line sublevel
stoping
4. Dip-line sublevel
stoping
III. Ground A. Mining with 1. Primary horizontal
control with alternating backfill and slicing with ascending
backfill and caving (descending) mining and
caving paste backfill and
secondary block
(sublevel) caving
2. Continuous room work
with primary mining with
backfill and secondary
mining with block
(sublevel) caving.
Thin and medium thick,
high dip orebodies.
Mining front divaricates.
Medium stable ores,
stable host rocks. Low
and medium grade ores.
Depth of application is to
be determined.
Thin and very thin, high
dip orebodies, medium
stable ores, stable host
rocks. Depth of
application is to be
determined.
Thick and very thick,
gently dipping and
weakly inclined
orebodies, unstable and
medium stable ores,
medium stable overlying
rocks. Medium grade ore.
Depth of application is to
be determined.
Thick and very thick,
gently dipping and
weakly inclined
orebodies. Medium stable
ores and host rocks.
Medium grade ore. Depth
of application is to be
determined.
Use of mobile machinery, High ore losses. High
high productivity. dilution, inclusive of
Automation. secondary dilution
(Figure 10).
Selective mining, without High labour intensity.
nonmetallic rocks, slot Great amount of
faces. Comparatively high preparatory driving. Use
ore extraction index. of portable equipment
(Figure 11).
Overlying rock mass Mining in stages. Higher
stability. Lower costs as requirements to ore
compared with Class I extraction standards.
methods. Lower extraction index as
compared with Class I
methods (Figure 12).
Overlying rock mass Mining in stages. Higher
stability. Higher stoping requirements to ore
productivity as compared extraction standards.
with the previous method. Lower extraction index as
compared with Method 1
in Class II.

6 7 Á-Á À Á À-À 2

3
7

1 2 5 8-12
4
À Á
FIG 1 - Class I, Group A, Type 1: 1 – inclined shaft; 2 – entry; 3 – levels; 4 – stoping; 5 – backfilling; 6 – level extraction;
7 – roof safety layer (please refer to Table 1).

8
Á-Á 7
1
2
À-À 4 3

À À

Á Á

5
1 3 6
FIG 2 - Class I, Group A, Type 2: 1 – inclined shaft; 2 – ore chute; 3 – cross adit; 4 – stopes; 5 – air raise; 6 – twin stope drifts;
7 – air-fill raise; 8 – filled levels (please refer to Table 1).

252 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CLASSIFICATION OF MINING METHODS FOR DEEP OREBODIES

1 4
3

FIG 3 - Class I, Group A, Type 3: 1 – stope; 2 – soft pillars; 3 – backfill; 4 – orebody (please refer to Table 1).

2 4

3,5
8-12 8-12
5 3 1
FIG 4 - Class I, Group A, Type 4: 1 – entry; 2 – extraction levels in a chamber; 3 – temporary pillar (secondary chamber);
4 – orebody; 5 – filling mass (please refer to Table 1).
61

3 5
61

2
61

FIG 5 - Class I, Group A, Type 5: 1 – gallery; 2 – ore chute; 3 – langwall; 4 – scraper excavator; 5 – pack; 6 – raise (please refer to Table 1).

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 253
V OPARIN, A TAPSIEV and A FREIDIN

À Á-Á 3
À-À Á 3 7

5
3,5-4,5

8-10
1 2 4 8
8-12 1
À 2 Á
FIG 6 - Class I, Group B, Type 6: 1 – haulage heading; 2 – drilling and haulage roadway; 3 – air ort; 4 – ventilation connection;
5 and 6 – air drifts; 7 – orebody; 8 – filling mass (please refer to Table 1).

6 À Á-Á 4 5
Á À-À

2
2

1 3
9 À 10 1 7 Á
FIG 7 - Class I, Group B, Type 7: 1 – haulage heading; 2 – drilling roadway; 3 – entry; 4 – air and backfill cross-drifts; 5 – air and backfill ort;
6 – filling holes; 7 – raise; 8 – broken ore; 9 – filled primary chamber; 10 – secondary chamber (pillar) (please refer to Table 1).

À-À Á-Á
Á À

2
2
40

3
8

Á 50 À

FIG 8 - Class I, Group A, Type 1: 1 – fan-patterned holes intended to cave safety shield; 2 – main fan-patterned holes; 3 – drilling and
haulage level; 4 – ore chute (please refer to Table 1).

254 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 CLASSIFICATION OF MINING METHODS FOR DEEP OREBODIES

À-À Á- Á
5

13-14 ì
2 3 3
Á 15-20 ì Á
1
4
12-14 ì

FIG 9a - Class II, Group B, Type 2: sublevel caving with areal-frontal ore drawing: 1 – haulage roadway; 2 – drilling and haulage cross-drift;
3 – loading slopes; 4 – broken ore; 5 – air drift (please refer to Table 1).

Á-Á À-À
À Á
22-30ì

3
11-15ì

10-15
À Á
40-60

FIG 9b - Class II, Group B, Type 2: sublevel caving with frontal ore drawing: 1 – ore chute; 2 – incline shaft; 3 – drilling and haulage
cross-drift (please refer to Table 1).

A B

FIG 10 - Class II, Group B, Type 3: (A) mid-thick orebody; (B) thin orebody: 1 – haulage heading; 2 – subdrift; 3 – holes; 4 – under cantilever
area; 5 – crown pillar; 6 – caved rock; 7 – broken ore (please refer to Table 1).

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 255
V OPARIN, A TAPSIEV and A FREIDIN

REFERENCES
Agoshkov, M I, 1965. Calculating and Designing Ore Mining Methods
and Technologies (Nauka: Moscow).
Baikonurov, O A, 1969. Classification and Selection of the Underground
Mining Methods (Nauka: Alma-Ata).
Bronnikov, D M, Zamesov, N F and Bogdanov, G I, 1982. Deep Ore
Mining (Nedra: Moscow).
Freidin, A M, Neverov, S A and Neverov, A A, 2008. Mine stability with
application of sublevel caving schemes, Journal of Mining Science,
44:82-91.
Imenitov, V R, 1978. Underground Mining Methods for Ore Deposits
(Nedra: Moscow).
Khomyakov, V I, 1984. Foreign Experience of Mining with Backfill
(Nedra: Moscow).
Oparin, V N, Rusin, E P and Tapsiev, A P, 2007. World Experience
Gained in Underground Mining Automation (SO RAN: Novosibirsk).
Pilenkov, Yu Yu, Freidin, A M and Antipov, A V, 1990. Estimation of
natural impact of the origination of rockburst hazard at shallow
FIG 11 - Class II, Group B, Type 4: 1 – haulage heading; depths, in Rock Pressure and Underground Ore Mining Technology
2 – inclined shaft; 3 – ore chute; 4 – haulage cross-drift for Large Depths (IPKON AN SSSR: Moscow).
(please refer to Table 1). Trushkov, N I, 1947. Ore Mining, Volume II: Mining Methods
(Metallurgizdat: Moscow).
Yun, R B, Gerasimenko, V I and Malyshev, V N, 1997. Dynamic rock
Backfill
pressure manifestations at Zheskazgan deposit, Gorny Zhurnal, No 3.

  
Orebody Zamesov, N F, 1979. Influence of the ground control on the design of the
mining method for gentle ore formations, in Problems of
Underground Ore Mining at Large Depths (IPKON AN SSSR:
Moscow).



 
Caved
 

FIG 12 - Combined ground control; caving and backfill

(please refer to Table 1).

256 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I

Resource Risk Quantification for a Deep Level Gold Mining
Operation
T Flitton1 and R Peattie2
ABSTRACT
South Africa’s Moab Khotsong Mine, one of the youngest mines to have
been brought into production in the historic Witwatersrand Basin, is
currently in its development phase, building up to full production around
2010. It is located in a structurally complex and highly faulted piece of
ground, with the bulk of the known mineral resource lying between 2100
and 3700 m below surface. Moab Khotsong is wholly owned by
AngloGold Ashanti Limited and has a published mineral resource of
13.79 million ounces of gold and reserves of 6.97 million ounces. In
addition to the gold resource, 39 million pounds of uranium exist within
the lease at a grade of 0.76 kg/t. The recent extension of Moab
Khotsong’s mining life has received much attention, with the potential for
the mine to operate until 2045. This extension is acutely strategic for both
the mine and the company, since the life of the operation has increased
significantly over the last few years as the exploration drilling continues
and the geological model evolves. A resource risk quantification exercise
was undertaken in early 2008 to account for the sensitivity of Project
Zaaiplaats 2 to the risks identified in an independent technical review.
This resource risk quantification exercise involved several scenarios
which tested the sensitivity of gold values – and thus gold content – to
changes in the geological facies and structural modelling. This paper will
briefly introduce the process followed for quantifying uncertainty in the
geological facies and the structural modelling, then present the results of
this exercise. The paper will highlight the potential financial impact that
this quantified change could have on the operation and demonstrate how
the geological risk has been incorporated into the financial modelling. A
summary and comparison of the results for each scenario will be
provided, as well as suggestions for further understanding the resource
uncertainty and quantifying its potential impact on the project.
INTRODUCTION
The AngloGold Ashanti South African operations comprise
seven underground mines which are located in two geographical
regions in the Witwatersrand Supergroup (deposited in the area
often described as the Witwatersrand Basin) – known as the
Carletonville (West Wits operations) and Klerksdorp (Vaal River
operations) gold fields (Figure 1 – left) (Anglogold Ashanti
Annual Report, 2007). Moab Khotsong Mine, the youngest of
AngloGold Ashanti’s South African operations (Figure 1), began
commercial production in January 2006 and is planned to be at
full production around 2010. The mineral resource at Moab
Khotsong is structurally complex and highly faulted, with large
fault-loss areas and the bulk of the known mineral resource lying
between 2100 and 3700 m below surface (AngloGold Ashanti
Annual Report, 2007). It is situated south and south-east of Great
Noligwa and Kopanang in the Klerksdorp (Vaal River) gold field,
in the Free State Province (AngloGold Ashanti Annual Report,
2007). The principal reef is the Vaal Reef, with mining grades of
between 10 g/t and 20 g/t gold. The Vaal Reef is commonly
represented by a series of oligomictic conglomerates and
quartzite packages developed on successive non-conformities.
Several distinct facies have been identified (Figure 2), each with
unique geological characteristics, gold distributions and grade
characteristics.
1. Geoscience Manager, AngloGold Ashanti Limited, Private Bag
X2051, Carletonville 2500, South Africa.
Email: Tflitton@anglogoldashanti.com
2. MAusIMM, Vice President, Mineral Resources Management,
AngloGold Ashanti Limited, PO Box 62117, Mashaltown 2107,
South Africa. Email: RPeattie@anglogoldashanti.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
Moab Khotsong is wholly owned by AngloGold Ashanti and
has a published mineral resource of 13.79 million ounces of gold
and reserves of 6.97 million ounces. In addition to the gold
resource, 39 million pounds of uranium (17 723 contained
uranium tonnes) exist within the lease at a grade of 0.76 kg/t
(AngloGold Ashanti Annual Report, 2007). Project Zaaiplaats 2
(PZ2: the extension project) is aimed at optimally extracting the
deeper portion (lower mine) of the Vaal Reef at Moab Khotsong,
with the potential then existing at Moab Khotsong for the mine
to operate until 2045. Investigations into extracting this block of
ground have been underway for several years, but unfavourable
market conditions have delayed the initiation of mining in this
deeper block of ground (up to 3500 m below surface).
Subsequent improvements in the world markets, not the least of
which has been the strengthening gold price, have lead
AngloGold Ashanti to revisit the deepening of Moab Khotsong’s
operations. A prefeasibility study on the viability of the project
was recently completed and Project Zaaiplaats 2 is currently in
the feasibility phase. Although the prefeasibility study shows a
relatively low return on investment, Project Zaaiplaats 2 is being
viewed in terms of the potential this project has of increasing the
life of the entire Vaal River operations (AngloGold Ashanti
Annual Report, 2007; de Beer et al, 2007). As a result, the
AngloAmerican Group’s Anglo Technical Division was
requested by AngloGold Ashanti to conduct a review of the
Project Zaaiplaats 2 prefeasibility study (Faul et al, 2007). A
number of recommendations, which were to be implemented prior
to feasibility, were made by Anglo Technical Division to address
some elements of the identified high-risks regarding geology and
mineral resources. This paper summarises the sensitivity analysis
that was undertaken in early 2008 to account for the uncertainty
regarding geology and mineral resources identified in the
independent technical review – it ameliorates the risk involved
with palinspastic reconstruction and interpretation of the
geological facies.
GEOLOGICAL SETTING
AngloGold Ashanti’s South African operations are all located in
the Witwatersrand Basin, regarded as one of the greatest
metallogenic provinces of the world (Robb and Robb, 1998). The
Witwatersrand Supergroup (2.7 Ga to 2.9 Ga) comprises a 6 km
thick sequence of predominantly argillaceous and arenaceous
sediments that extend laterally for some 300 km north-east/south
-west and 100 km north-west/south-east on the Kaapvaal Craton.
The upper portion of the sequence (the Central Rand Group)
contains the laterally extensive, economically mineralised, gold-
bearing quartz pebble conglomerate horizons or reefs (Robb and
Robb, 1998). Only a small portion of the Witwatersrand Super-
group is exposed in outcrop, most being overlain, unconformably,
by rocks of the Venterdorp (2.7 Ga), Transvaal (2.6 Ga) and Karoo
(2.80 Ma) Supergroups (Robb and Robb, 1998; McCarthy, 2006).
The reefs, generally less than two metres thick, are widely
considered to represent laterally extensive braid-plain and lesser
alluvial fan environments (Robb and Robb, 1998). The highly
faulted Klerksdorp gold field (Figure 1) lies to the southwest of
the Caletonville gold field and is separated by the Potchefstroom
Gap. The geology of the area is complicated by erosional
truncation of lower formations and overstepping by younger
formations (Robb and Robb, 1998). There is still much debate
Spectrum Series Volume 17 257
T FLITTON and R PEATTIE

FIG 1 - Location map of the Witwatersrand Basin in South Africa, with younger cover sequences removed (left), showing the Vaal River
operations (right).

Moab Khotsong Mine

Great Noligwa Mine

Kopanang Mine

Facies interpretation
at pre-feasibility

Great Noligwa Mine Great Noligwa Mine

Kopanang Mine Kopanang Mine

Scenario D Scenario E
optimistic 500 pessimistic 500
Facies Facies

FIG 2 - Varying facies interpretations – facies interpretation at prefeasibility (top) versus scenario D (bottom left) and scenario E
(bottom right).

258 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 RESOURCE RISK QUANTIFICATION FOR A DEEP LEVEL GOLD MINING OPERATION
about the origin of the gold mineralisation in the Witwatersrand
Basin. Gold was generally considered to have been deposited
syngenetically with the conglomerates, but increasingly an
epigenetic theory of origin is being supported (Robb and Robb,
1998). Nonetheless, the mineralisation is invariably associated
with erosional disconformities within the sequence in various
host lithologies. Gold generally occurs with pyrite, uraninite and
carbon, with quartz being the main gangue mineral (Robb and
Robb, 1998).
PROJECT ZAAIPLAATS 2 MINERAL RESOURCE
Facies model
The facies model (Figure 2) within Project Zaaiplaats 2 area is
based largely on 15 surface boreholes (multiple reef intersections
per borehole) and prior knowledge from the Great Noligwa and
Kopanang Mines to the north (Figure 3). Experience has shown
that an understanding of the processes that led to the formation
of the reef and the distribution of gold within that orebody is of
crucial importance for estimation to be as accurate as possible
(de Beer et al, 2007). The Vaal Reef package on Great Noligwa,
Kopanang and Moab Khotsong Mines has been shown to consist
of three correlatable conglomerate units – the ‘A’, ‘B’ and ‘C’
members. Changes in the nature and inter-relationships between
these units are reflected in the broad-scale distribution of gold on
Great Noligwa and Kopanang. Such changes are often subtle and
require detailed logging and correlation work. The geological
characteristics that define the reef in each of the areal
subdivisions (geozones) are then compared to counterparts in
borehole core, and the ‘Facies Model’ is extended into
Kopanang Mine
Great Noligwa Mine
±1
50
0m
Great Noligwa Mine
Kopanang Mine
Scenario B
rec onstruction
±430m further west
FIG 3 - Varying palinspastic reconstruction – reconstruction at prefeasibility (top) versus scenario B (bottom left) and scenario C
(bottom right).
Advances in Orebody Modelling and Strategic Mine Planning I
poorly-explored areas (de Beer et al, 2007). Using the mapped
geological characteristics of the conglomerates of the A and C
horizons, the Kopanang-Great Noligwa-Moab Khotsong mining
area is currently subdivided into nine geozones, which are
viewed as geologically homogenous regions suitable for
estimation (de Beer et al, 2007). Only four of these geozones
(500/370/440/460) are present within the Moab Khotsong Lower
Mine (Figure 2). The 470 geozone currently present in Great
Noligwa Mine and Kopanang Mine between the 500 and 440
geozones is not present in the Moab Khotsong Lower Mine.
Structural model
The structural component of the geological model was
synthesised from the analysis of vibroseis survey data, along
with the structural information projected from contiguous mining
areas. Important stratigraphic and structural information was then
derived from the detailed analysis of the 15 surface boreholes and
their deflections that intersected the area, and used together with
vibroseis interpretation, to identify the major faults in the area.
These major faults were then used to define the margins of the
Vaal Reef blocks and the expected Vaal Reef horizon was then
constructed by projection parallel to the structural dip as far as
the major faults (de Beer et al, 2007). The resulting structural
architecture was tested by reconstruction. In this process, all reef
blocks are palinspastically restored to their original unfaulted
locations prior to movement on the Jersey Fault (movement
modelled to be 1500 m to the south-east), adjacent to the Great
Noligwa and Kopanang mines (Figure 3). By restoring the reef
blocks to their interpreted unfaulted locations, it becomes
possible to interpret how the different units of the Vaal Reef,
Moab Khotsong Mine
Reconstruction
at pre-feasibility
Kopanang Mine
Scenario C
reconstruction
±640m further east
Spectrum Series Volume 17 259
T FLITTON and R PEATTIE
intersected in the surface drilling from the Moab Khotsong
blocks, correspond with the projected facies plans from the
adjacent Great Noligwa and Kopanang properties (de Beer et al,
2007). In addition to the structural blocks thus identified,
extensive geological target areas are known to exist, but at a
poorly-defined elevation. Exploration drilling in these blocks is
likely to realise the upside potential of the project (shown in
Figure 3 as the hatched reef blocks). Geological target areas do
not form part of the published mineral resource (de Beer et al,
2007).
Estimation methodology
The sampling data for Project Zaaiplaats is comprised of
underground chip samples, underground boreholes and surface
boreholes, with the surface boreholes being of a different support
to the underground boreholes and chip sampling. All sample
locations are reported as a composite over a mineralised width,
resulting in a single channel width (cm) and metal accumulation
(cm g/t) value.
AngloGold Ashanti makes use of a Bayesian geostatistical
approach for resource estimation at project Zaaiplaats 2, where the
absence of dense sampling data gold estimation is based on a
combination of the observed data and external knowledge relating
to the data. A Bayesian geostatistical approach asserts that the area
to be evaluated forms part of a larger continuous entity, to which
the observed data belongs. The assumption behind this approach is
that the geological, statistical and spatial characteristics observed
in the known areas hold for the ‘unknown’ uninformed areas. If
the assumption is valid, then the orebody can be evaluated over
large uninformed areas, where limited information exists. Geozone
constraints are critically important within this process, because the
geozones subdivide the data into distinct populations and the
parameters of these populations play a critical role in the
development of the estimates (Chamberlain, 1997).
Mixed support cokriging was used in the estimation of the
mineral resource for the project. It is a technique that enables the
use of data of mixed support, allowing both borehole and
underground sampling data to be used together (Chamberlain,
1997). Project Zaaiplaats is currently estimated as part of a larger
entity, recognising that the Vaal Reef is present over a number of
mines. The models that are developed are therefore on a regional
scale, rather than a mine scale. Estimation is performed into 300 m
× 300 m sized blocks, which fully capture the within block
variance, allowing the cokriging of data of different support sizes
over long ranges. The variography is based on the combination of
the block variograms and borehole variograms (Chamberlain,
1997). Estimation is done in log space because of the highly
skewed gold distribution found within the Vaal Reef. The final
gold estimates are then calculated by back transforming the kriged
estimates, using the compound log normal (CLN) four parameter
distribution models (Chamberlain, 1997). The current method has
proven reliable (resource estimates derived from mixed support
cokriging have been compared against production results on well
established mines in the same area) and has been in use on the
Witwatersrand reefs, particularly the Vaal Reef, for a number of
years. The mixed support cokriging global estimation method has
undergone intense scrutiny by a number of external auditors,
including AngloAmerican Group’s Anglo Technical Division,
Snowden and SRK.
Due to the limited amount of borehole data within the project
area (15 mother surface boreholes) and the distance of the project
to surrounding mines (less than one kilometre), the Moab
Khotsong Lower Mine reef blocks have been reconstructed back
to their original unfaulted positions, prior to movement on the
Jersey Fault, adjacent to Great Noligwa and Kopanang mines
(Figure 3). There is an implicit risk in this, in that the ‘true’
movement vectors could be different to what is currently used, but
the benefit in terms of additional sampling information being
available potentially outweighs the risk. Any estimation process is
largely reliant on the validity of the data as well as the accuracy
260 Spectrum Series Volume 17
and relevance of the geological model. This is particularly the case
in a Bayesian geostatistical approach, and even more so when
reconstruction of data has taken place as in project Zaaiplaats.
RESOURCE RISK QUANTIFICATION
The structural modelling (palinspastic reconstruction) and the
geological facies were identified as high-risks in the Independent
Technical Review Report (Faul et al, 2007). They suggested the
following course of action for risk mitigation:
The reconstruction of the blocks to a pre-Jersey
Fault position needs to be tested as it is indicated
that there is an alternative to the current model
with a smaller sinistral component.
The facies boundaries used for the current
reconstruction of a pre-Jersey fault situation are
soft boundaries, transitions rather than hard
boundaries. Therefore the interpretation of the
facies in the main eastern block of the project
needs to be tested with borehole intersections as
the knowledge of the possible presence or absence
of the barren 470 facies is of economic
importance. In the current model the 470 facies
does not occur in the project area, however, it
may still occur in the place of the current
440 facies.
The resource risk quantification exercise was undertaken to
ameliorate the risk involved with palinspastic reconstruction and
interpretation of the geological facies. The resource risk
quantification exercise involved the following scenarios, testing
the sensitivity of values, ie gold, to changes in the geological
facies and reconstruction:
A. Moab Khotsong Lower and Upper Mine data in
unconstructed space, rekriged and re-estimated.
B. Moab Khotsong Lower Mine reconstruction further west
compared to reconstruction used in prefeasibility, rekriged
and re-estimated.
C. Moab Khotsong Lower Mine reconstruction further east
compared to reconstruction used in prefeasibility, rekriged
and re-estimated PZ2 with more optimistic view on 500
facies.
D. Moab Khotsong Lower Mine geological facies having an
optimistic outlook on 500 facies.
E. Moab Khotsong Lower Mine geological facies having a
pessimistic outlook on 500 facies.
F. Moab Khotsong Lower and Upper, Great Noligwa Mine
and Kopanang Mines having no facies boundaries (data
combined), rekriged and re-estimated.
G. Different global estimation methods.
Scenario A
Scenario A is intended to quantify the sensitivity on Project
Zaaiplaats 2 if the faulted blocks were not reconstructed at all.
The data (samples from boreholes and/or chip samples) as well
as the structure would remain in true space for Moab Khotsong
Upper and Lower Mines (Figure 3).
Scenario B
Scenario B is intended to quantify the sensitivity on Project
Zaaiplaats 2 if the faulted blocks were reconstructed a further
430 m west from the reconstruction used in the prefeasibility
(movement modelled to be 1500 m to the south-east). The data
(samples from boreholes and/or chip samples), as well as the
Advances in Orebody Modelling and Strategic Mine Planning I
 RESOURCE RISK QUANTIFICATION FOR A DEEP LEVEL GOLD MINING OPERATION

TABLE 1
Comparison of results from the resource risk quantification exercise.

Project Zaaiplaats 2 Au (Moz) Diff in Au (Moz) % diff in Au (Mz)

Feasibility 7.683
Scen A – No palinspastic reconstruction 6.932 -0.751 -10.84%
Scen B – Palinspastic reconstruction further west 7.090 -0.593 -8.37%
Scen C – Palinspastic reconstruction further east 7.566 -0.117 -1.54%
Scen D – Optimistic 500 facies 7.796 0.113 1.44%
Scen E – Pessinistic 500 facies 7.525 -0.158 -2.10%
Scen F – No facies boundary (max samples 30) 8.033 0.350 4.35%
Scen G – Average value 6.187 -1.496 -24.18%
Scen G – CLN estimate using chip samples variances 7.694 0.011 0.14%
Scen G – CLN estimate using surface borehole variances 6.390 -1.293 -20.23%
Scen G – Sichel – T estimate using chip samples variances 8.139 0.456 5.60%
Scen G – Sichel – T estimate using surface borehole variances 6.378 -1.305 -20.45%
Scen G – Two parameter lognormal estimate 6.749 -0.934 -13.84%

structure of the Moab Khotsong Lower Mine faulted blocks,
would be reconstructed a further 430 m west of the re-
construction used at prefeasibility. Moab Khotsong Upper Mine
would maintain its current interpretation of reconstruction as
used in the prefeasibility, as it is a different faulting system
compared to the Jersey Fault (Figure 3).
Scenario C
Scenario C, like scenario B, is intended to quantify the sensitivity
on Project Zaaiplaats 2 if the faulted blocks were reconstructed a
further 640 m east of the reconstruction used during prefeasibility.
The data (samples from boreholes and/or chip samples), as well as
the structure of the Moab Khotsong Lower Mine faulted blocks,
would be reconstructed a further 640 m east of the reconstruction
used at prefeasibility. As with scenario B, all the data as well as
the structure of the Moab Khotsong Upper Mine faulted blocks,
would only be reconstructed using the reconstruction vectors at
prefeasibility (Figure 3).
Scenario D
Scenario D is intended to quantify the sensitivity on Project
Zaaiplaats 2 if the facies lines were adjusted, whilst still keeping
the true facies interpretation of the surface boreholes. A more
optimistic interpretation of the 500 reef facies is used in place of
the facies interpretation at prefeasibility (Figure 2). All the data
(samples from boreholes and/or chip samples), as well as the
structure of the Moab Khotsong Upper and Lower Mine faulted
blocks, would be reconstructed using the reconstruction vectors at
prefeasibility.
Scenario E
Scenario E, like scenario D, is intended to quantify the sensitivity
on Project Zaaiplaats 2 if the facies lines were adjusted, whilst still
keeping the true facies interpretation of the surface boreholes. A
more pessimistic interpretation of the 500 reef facies is used in
place of the facies interpretation at prefeasibility (Figure 2). All
the data (samples from boreholes and/or chip samples), as well as
the structure of the Moab Khotsong Upper and Lower Mine
faulted blocks, would be reconstructed using the reconstruction
vectors at prefeasibility.
Scenario F
Scenario F is intended to quantify the sensitivity on Project
Zaaiplaats 2 if no facies lines (ie hard boundaries and similar
geological facies characteristics are ignored) are used when
doing the estimation. This scenario will address the absence of
the 470 reef facies at Moab Khotsong Lower Mine. All data is
merged (including the 470 facies) into one single dataset and
estimated accordingly. All the data (samples from boreholes and/
or chip samples), as well as the structure of the Moab Khotsong
Upper and Lower Mine faulted blocks, would be reconstructed
using the reconstruction vectors at prefeasibility.
Scenario G
Various non-spatial global estimation methods were used to
calculate global estimates for various portions of the Moab
Khotsong Lower Mine, thus obtaining a range of values for the
Project Zaaiplaats 2 area. The methods included arithmetic
average of the surface boreholes, a two parameter lognormal
estimate using value of surface boreholes, Sichel-T estimates
using surface boreholes, CLN estimates using surface boreholes.
All of these estimates are further compared to the current
prefeasibility estimates for Project Zaaiplaats 2.
SUMMARY AND COMPARISON OF RESULTS
It was noted that for scenarios A, B and C, where reconstruction
movements were varied (Figure 3), the new mineral resource
estimates all decreased (Table 1). In the scenario where no re-
construction was used (A) the highest difference in value
compared to the prefeasibility mineral resource was calculated.
Scenario A is not considered as a realistic alternative based on the
current understanding of the geology, however, it does give a
worst-case scenario if either the reconstruction vectors utilised are
inaccurate, or if the mineralisation was post the final fault
movement. Scenario B is therefore considered as the most likely
worst-case scenario (highest change in value from prefeasibility) –
the potential financial impact this quantified change could have on
the operation is shown in Table 2.
Scenarios D and E, where facies interpretation was varied, while
honouring the surface boreholes logged facies interpretation as
used in the prefeasibility, showed negligible difference compared
to the prefeasibility mineral resource estimate (Table 1). While for
scenario F, where no geological modelling was used in the
estimation, the estimates were higher than those calculated for the
prefeasibility. The relationship of variance for the combined data,
compared to each facies individually, resulted in more weighting
given to closer (higher value) regularised blocks in this scenario.
Scenario G involved the calculation using various global
estimation methods, resulting in a range of values for various

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 261
T FLITTON and R PEATTIE
TABLE 2
Effect on IRR (standardised) of PZ2 if scenario B (worst-case
scenario – further west palinspastic reconstruction) is used as the
likely reconstruction.
Prefeasibility Scenario B Percentage
difference
On reef value (cm g/t) 2597 2401 -8%
Standardised IRR (%) 1 0.75 -25%
portions of the Moab Khotsong Lower Mine. The majority of the
global estimates calculated in scenario G are below the
prefeasibility Project Zaaiplaats 2 published value estimate. This
scenario highlights that Project Zaaiplaats 2 is very sensitive to the
amount of the project area that is covered by the 370 facies.
Project Zaaiplaats 2 has a high published resource value estimate,
mainly because the bulk of the ground consists of the ‘highest
value’ 370 reef facies. Thus a very real risk to the extension
project exists if there is not as much 370 reef facies as is currently
modelled, in which case the value can only decrease.
RECOMMENDATIONS FOR FURTHER STUDY
Two additional scenarios are to be completed in the resource risk
quantification exercise. A multi-gaussian conditional sequential
simulation is suggested as a means to get a range of estimates
located in space for the project area. A final scenario where the
470 facies is shown to be present in the Moab Khotsong Lower
Mine, using the data from the 470 facies in the north to predict the
262 Spectrum Series Volume 17
values to the south (together with the reconstruction vectors at pre-
feasibility of 1500 m north-west from faulted true positions),
should be completed to potentially quantify the lack of presence of
this facies. Both scenarios are to be completed in the near future,
thus fully quantifying the sensitivity on Project Zaaiplaats 2.
REFERENCES
AngloGold Ashanti Annual Report, 2007. Supplementary information:
Mineral resources and ore reserves, p 124.
Chamberlain, V A, 1997. The application of macro co-kriging and
compound lognormal theory to long range grade forecasts for the
Carbon Leader Reef, project report submitted in partial fulfillment of
the requirements of Master of Science in Engineering (unpublished),
University of the Witwatersrand.
De Beer, R, McIvor, J, Broodryk, A, Kotze, B, Uys, J, Engelbrecht, R,
Flitton, T M, Frith, I, Botha, L and Charalambous, V, 2007. Africa
underground region, Moab Khotsong Mine, business plan 2008: R3,
Bp 15 Book, Project Zaaiplaats 2, study (unpublished).
Faul, H, Rompel, A K K, Rice, P, Carvil, P, Roman, N, Meyer, G,
Stephenson, J D, van der Linde, A, Hacking, T, Hutton, C, Johnson,
A, Magqabi, L, Oldfield, R and Cantrell, K, 2007. Independent
technical review report on behalf of the AAplc group technical
director of the AngloGold Ashanti Ltd Moab Khotsong Mine PZ2
Zaaiplaats prefeasibility study (unpublished).
McCarthy, T S, 2006. The Witwatersrand Supergroup, in The Geology of
South Africa (eds: M R Johnson, C R Anhouser and R J Thomas),
pp 155-186 (Geological Society of South Africa, Council for
Geoscience: Pretoria).
Robb, L J and Robb, V M, 1998. Gold in the Witwatersrand Basin, in The
Mineral Resources of South Africa Handbook (eds: M G C Wilson
and C R Anhouser), 16:294-349 (Council for Geoscience).
Advances in Orebody Modelling and Strategic Mine Planning I

An Application of Genetic Algorithms to the Main Ventilation
System Optimisation for Multi-Level Mine Operations –
The Impact of the Fitness Function
E Acuña1, S Hardcastle2,3, L Fava4,5, P Dunn6 and S Hall7
ABSTRACT
An operational ventilation network has to deliver the airflow required for
the planned activities in an underground environment. The objective in
designing a ventilation network is to select the fan locations and
operation points that will provide the required pressures and airflows
while achieving minimum or near-minimum cost. Historically, the
methodology for ventilation optimisation has been based on case studies
and expert knowledge. Additionally, some more rigorous mathematical
approaches have been developed, and applied to different networks. The
optimisation of primary mine ventilation systems using genetic
algorithms has been demonstrated under two approaches: with a fitness
function based on airpower consumption, and with a fitness function
based on the deviation of airflow from the requirements. Unfortunately,
the different approaches were applied to different networks, making it
difficult to compare the benefits and tradeoffs of the fitness functions.
The main objective of the proposed approaches is to improve the value of
mining plans by accounting for ventilation requirements, where the
improvement can be realised by two means: first, by allowing the
schedules to be executed as planned, reducing (or avoiding) development
and production losses, and second, by reducing ventilation costs. This
paper presents a comparison between the airpower and the airflow
deviation fitness functions to improve main mine ventilation systems. An
application to a mine ventilation network is presented and solved with
both fitness functions, and then compared to determine the advantages
and disadvantages taking into account the airpower and airflow deviation
values obtained for the network and the solving time.
INTRODUCTION
An adequate mine ventilation system is one of the key
components to allow development and production in
underground environments. Generally, the mine ventilation
system can be divided into two levels or functions: primary and
secondary. The primary (or main) ventilation system, containing
the main and booster fans, is solely responsible for generating an
1. PhD Candidate, Natural Resources Engineering, MIRARCO,
Laurentian University, 935 Ramsey Lake Road, Sudbury Ontario
P3E2C6, Canada. Email: eacuna@mirarco.org
2. Adjunct Professor, School of Engineering, Laurentian University,
935 Ramsey Lake Road, Sudbury Ontario P3E2C6, Canada.
Email: shardcas@nrcan.gc.ca
3. Senior Research Scientist – Ventilation Specialist, Mine Air Quality
and Ventilation, CANMET, Natural Resources Canada, 1079 Kelly
Lake Road, Sudbury Ontario P3E5P5, Canada.
Email: shardcas@nrcan.gc.ca
4. Adjunct Professor, Department of Mathematics and Computer
Science, Laurentian University, 935 Ramsey Lake Road, Sudbury
Ontario P3E2C6, Canada. Email: lfava@symboticware.com
5. Research and Development Manager, Symboticware Incorporated,
83 Durham Street, Sudbury Ontario P3E3M5, Canada.
Email: lfava@symboticware.com
6. Director – Western Australian School of Mines, Faculty of Science
and Engineering, Curtin University of Technology, GPO Box U1987,
Perth WA 6000. Email: P.Dunn@curtin.edu.au
7. President and CEO/Professor, School of Engineering, MIRARCO,
Laurentian University, 935 Ramsey Lake Road, Sudbury Ontario
P3E2C6, Canada. Email: shall@mirarco.org
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
airflow from the surface to the vicinity of mining areas, then
back to the surface. The secondary (or auxiliary) ventilation
system is responsible for creating an airflow from the primary
system to the dead-ends where mining development and
production generally occur. The main purposes of ventilation are
to bring fresh air (oxygen) to the different areas in the
underground environment for people and machines, then to dilute
and remove contaminants generated by the mining process.
Adequate ventilation is one of the key conditions enabling the
mine to safely perform the activities scheduled by mine planners.
Considering the operation of fans, along with heating and
cooling systems to condition the air, ventilation can account for 15
to 22 per cent of a mine’s operating cost, as stated by Petrov and
Popov (2004). However, from an energy usage perspective,
ventilation costs in mechanised mines can account for around
40 per cent of the consumption, according to Hardcastle and
Kocsis (2007); the proportion of energy usage by primary,
secondary, heating and cooling can vary considerably as stated by
Hardcastle et al (2007). Consequently, the energy consumption of
ventilation systems is becoming one of the most important cost
control areas within underground mines. With increasing pressure
on mines to reduce energy consumption and greenhouse gas
(GHG) emissions, mining operations are being forced to develop
new approaches and design tools, and also to incorporate the latest
mining technologies in order to lower production costs while
meeting production and environmental requirements, according to
Kocsis, Hall and Hardcastle (2003).
Currently, the optimisation of mine ventilation systems is
typically a manual decision process. It is assisted by computers,
using commercial ventilation distribution solvers (referred to as
network simulators within the mining industry) to test alternatives
proposed by the ventilation expert. From those alternatives, the
one that reports the lowest energy consumption value while
satisfying the peak airflow requirements is typically selected and
called the optimal solution. Mine ventilation optimisation based
on ventilation solvers is accepted by the mining industry as a
reliable methodology to decide good feasible alternatives to
implement in mine ventilation networks. The decision as to
whether or not such an optimised solution is indeed the optimal
solution is dependent on the input criteria, and how well it
represents the actual conditions within the mine over time.
Because ventilation requirements can fluctuate significantly in the
short term, while overall demand changes gradually over the long
term, these manually selected solutions could be suboptimal. To
date, various attempts have been made (Lowndes, Fogarty and
Yang, 2005; Lowndes and Yang, 2004; Calizaya, McPherson
and Mousset-Jones, 1987; Wu and Topuz, 1998) to generate a
mathematical mine ventilation optimisation technique or process
and have concentrated on a single time period and its optimal fan
selection. Aside from Lowndes and Yang (2004) and Acuña et al
(2009), mine ventilation optimisation techniques have not been
implemented across multiple periods.
BACKGROUND
Mine ventilation network optimisation
Mine ventilation networks are modelled as sets of connected
nodes and edges – these are respectively referred to as junctions
Spectrum Series Volume 17 263
E ACUÑA et al
and branches by mine ventilation design practitioners. The edges
represent the different paths that the airflow can follow; the
nodes are the connections between those paths. Each edge has an
associated resistance value that represents the friction that has to
be overcome by the airflow in order to traverse the edge. Because
of this friction, a pressure drop is generated in every branch
traversed by airflow. Fans are used to supply the pressure in
order for the airflow to overcome the pressure drop of the
different branches, and to deliver the right amount of airflow
through the ventilation network. Mine ventilation network
calculations are based on the network designs and Kirchhoff’s
first and second laws, as stated by Lowndes, Fogarty and
Yang (2005) and McPherson (1993). Mine ventilation network
optimisation problems are nonlinear and nonconvex according to
Xing and Topuz (1998). This means that no known classical
optimisation method can solve this problem and guarantee an
optimum. Moreover, there is no known function to transform the
pressure distribution (vector P) into airflows in the mine
ventilation network. Iterative or optimisation methods have to be
used to generate the airflow distribution. In addition, depending
on the problem, the solution space for single and multiple period
network problems can be large enough to make it impossible to
find the optimum solution in a reasonable time.
Mine ventilation network optimisation in terms of energy is
written as a nonlinear optimisation problem of the following
form:
b
Pj ⋅ Qj
Min f ( x ) = ∑ 24 ⋅ Days ⋅ ⋅ EC
j=1 nj ⋅1000 (1)
P = ( P1 ,..., Pb )t ∈ F ⊆ S ⊆ R b
where:
b is the number of branches in the network
Days is the number of days that the fan is going to be working,
usually on a yearly basis, ie, Days = 365; 24 is used to
transform the days into hours; EC is the energy cost
($/kWh); and 1000 transforms W into kW because of the
EC units. P = (P1, … , Pb) is the vector of pressures, Q =
(Q1, … , Qb) is the vector of airflow resulting from the
pressures, ç = (ç 1, … , ç b) is the vector of efficiencies for a
combination of pressure and airflow if a fan is located in
the branch j; F is the feasible region, and S is the whole
search space
subject to:
b Lowndes and Yang, 2004). The most popular technique used to
∑a ij ⋅Qj = 0 ∀i = 2,..., n
j=1
where:
Qj is the airflow quantity through branch j, n is the number of
nodes in the network, aij = 1 if branch j is connected to node i and
the airflow goes away from node i, aij = -1 if branch j is
connected to node i and the airflow goes into node i, aij = 0 if
branch j is not connected to node i;
b
∑b ij ⋅ H j = 0 ∀i = 2,..., m
j=1
H j = HL j + HR j − HFj − HN j ∀j ∈ {1,... b}
HLj = Rj ⋅ Q 2
j ∀j ∈ {1,..., b}
where:
HLj is the friction pressure drop for branch j
264 Spectrum Series Volume 17
Rj is the resistance factor for branch j
HRj is the pressure drop of the regulator in branch j
HFj is the fan pressure in branch j
HNj is the natural pressure in branch j
m is the number of chords (meshes) in the network (m = b – n
+ 1), bij = 1 if branch j is contained in mesh i and has the
same direction, bij = -1 if branch j is contained in mesh i
and has the opposite direction, and bij = 0 if branch j is not
contained in mesh i
L j ≤ Qj ≤ U j ∀j ∈ {1,..., b} (6)
HR j ≥ 0 ∀j ∈ {1,..., b} (7)
where:
Lj and Uj are the lower and upper bound for the airflow in each
branch of the network and HRj is the additional resistance that
can be installed in each branch.
The objective function and the pressure balance constraints
(Kirchhoff’s second law) are nonlinear. The airflow balance
constraints (Kirchhoff’s first law) are linear constraints. Airflow
requirements as well as lower and upper bounds for the airflow
in the different branches can be added. A vector P that satisfies
all the constraints is a feasible solution of the problem. The
feasible region F is the set of all feasible points. The objective of
the optimisation process is to find the best feasible P* ∈ F such
that f(P*) = f(P) for all P ∈ F. Vector P* is called the optimum
solution of the problem. In practical terms, the vector P* is the
set of pressure values for the fans that will deliver at least the
required airflow for the mine, and at minimum cost.
In this study, two different fitness functions are implemented
and compared: one based on the airpower and the other one on
the square deviations from the airflow requirements. The idea is
to study the capability of these fitness functions to explore the
solution space and find good feasible solutions or near optimal
solutions. The principal assumption is that a solution with
reduced airflow volumes will be less expensive in terms of
airpower or energy cost, so it is expected that both fitness
functions should converge to the same optimal or near optimal
values.
In order to solve these equations, nonlinear optimisation or
iterative methods have been successfully implemented (Wu and
Topuz, 1998; Calizaya, McPherson and Mousset-Jones, 1987;
(2) solve mine ventilation networks, ie, find the airflows of the
network based on the pressures of the different fans, is the Hardy
Cross method. This method is based on an initial solution for the
ventilation network, and then converges to the real values of the
pressure drop and airflow on each of the branches of the network
through an iterative process. The method is popular because for
most of the ventilation networks, it converges in a small number
of iterations, as stated by McPherson (1993). It uses a
simplification of Kirchhoff’s first law; the method assumes the
densities to be the same in all junctions. This is called the
incompressible approach, and the mass flow conservation is
transformed into airflow conservation. The incompressible
(3)
approach is only valid for shallow mines to a depth of 500 m. For
deeper mines, the compressible approach is recommended.
(4)
Practical mine ventilation network optimisation
(5) Several mine ventilation solvers are available on the market as
software tools. They are called mine ventilation simulation tools,
although they don’t perform any simulation process in the sense
of random variable distributions and discrete or continuous event
simulation. These tools solve the network and calculate the
Advances in Orebody Modelling and Strategic Mine Planning I
 AN APPLICATION OF GENETIC ALGORITHMS TO THE MAIN VENTILATION SYSTEM OPTIMISATION
values of the airflow in the ventilation networks. This is why
they are known as ventilation network simulators, because they
are able to ‘simulate’ (calculate the value of) the airflows. These
software tools have a network solver connected to a user
interface to display the network and the analyses available.
In order to solve mine ventilation networks, the Hardy Cross
technique is used to calculate the airflows in the network, based
on the locations of the fans in the branches, resistances of the
branches, and the duties of those fans (fixed pressure or
pressure-airflow curves). The ventilation engineer can manually
interact with the user interface to change the network parameters,
or the fan locations and duties, and generate a new result for each
scenario, but the solver will only generate the solution for that
particular scenario. Ventilation solvers have no capacity to
propose fan locations and duties in order to have a feasible
solution, or the best fan locations and duties to meet the airflow
requirements and generate a feasible, economical solution. Based
on the trial of several different combinations of the fan locations
and duties, using mine ventilation solvers (with the Hardy Cross
technique), the best solution is identified and taken as optimal.
Constrained genetic algorithms
Constrained genetic algorithms have been successfully used to
solve one period, large mine ventilation network problems, as
stated by Lowndes, Fogarty and Yang (2005) and Acuña et al
(2009), using a penalty function with a large preset value to
preserve feasibility. A constrained genetic algorithm doesn’t
explore the entire solution space (S) with optimality as its prime
target, then project to the feasible region; instead, it explores the
space of feasible solutions (F) to find the best feasible solution
possible. For this reason, the use of penalty functions can
increase the likelihood that the final solution of the genetic
algorithm will be just a local optimum. Also, depending on the
solution space and the operators of the genetic algorithm, the
final solution can be infeasible, although this outcome may be
unlikely. Considering the financial relationship between
production, development and ventilation, the foremost goal will
always be to have a feasible solution for ventilation to support
the planned schedule, though the additional benefit of an
improved solution will certainly be well received.
GENETIC ALGORITHM IMPLEMENTATION
Genetic algorithms (GAs) are search algorithms inspired by
Darwin’s theory of evolution. To apply a GA, an initial
population of individuals is created at random, with each
individual representing a solution to the problem. Next, pairs of
individuals are combined by means of a crossover operation to
produce offspring for the next generation. This leads to an
intense search in one area of the solution space. In order to
explore more areas of the solution space, a mutation process is
also used to randomly modify some of the individuals of each
generation. The process is repeated until the stopping criterion
is reached, and the best solution found is retained. The
probability of crossover for an individual is proportional to its
fitness; the fitness of an individual is a measure of the quality of
the solution. The use of a genetic algorithm leads to an
improvement of the individuals in successive generations,
according to Bodenhofer (2007) and Goldberg (1989). The
pseudocode of the GA used in this paper is given as follows.
Begin
Compute initial random population V (size n)
Evaluate individuals of V with 3D-CANVENT ventilation
solver
Compute fitness of individuals of V
While Termination criterion is not reached do
Select pairs of individuals from V for crossover
Advances in Orebody Modelling and Strategic Mine Planning I
Generate n offspring V’ by crossover of selected pairs and
mutation
(Crossover rate 0.6 and Mutation rate 0.1)
Evaluate individuals of V’ with 3D-CANVENT ventilation
solver
Calculate feasibility value, penalty and fitness of V‘
Let V = n best individuals from V and V’
Return best solution (individual) found, and feasibility value
End
CASE STUDY
The case study is based on the mine ventilation network of the
former Gedling Colliery coal mine in the United Kingdom. The
mine has one fresh air intake shaft and one exhaust shaft, with a
main surface exhaust ventilation fan system (MF), as stated by
Hardcastle (1983). A simplified schematic of the ventilation
network is displayed in Figure 1. There are five work faces or
production areas, labelled WF1 through WF5 on the diagram;
WF1 and WF5 are simplifications of more extensive parts of the
ventilation system with multiple work faces. The airflow
requirements are specified for each area, and they depend on the
type of activities taking place. In addition to the main surface fan,
there are two booster fan systems (F1 and F2) required to assist
the surface fan duties in order to provide the airflow requirements
through the mine. An additional booster fan is located in the
network to simulate the natural ventilation pressure.
There are three main objectives assigned to the GA decision:
the number of fans to install in the mine, the fixed pressure of the
fans (that will determine the airflows generated by those fans,
usually duties) and the location of the fans given the available
places within the mine ventilation network. The final decision
has then three parts, ie (1) number (2) duties and (3) locations,
for the network. The decision is the value of the fixed pressure to
be applied in a given location in the network. The number of
locations with a fixed non-zero pressure value is the number of
fans that will be needed in the mine (decision 1). The duties are
the combination of the fixed pressures and the airflow resulting
from the fixed pressure combination (decision 2). Finally, the
locations will be those with non-zero pressure values (decision 3).
This study does not claim that the GA will find the optimal
solution to the mine ventilation network optimisation; in fact, it
is recognised that GAs cannot provide a certificate of global
optimality. For each run of the algorithm, the best solution found
by the GA will be displayed; the term ‘optimal solution’ is
avoided here. The feasibility of the solutions found by the GA
will be discussed as well.
Chromosome representation
The chromosome representation (string) is intended to represent
the solution vector of the problem, and also handle as many
constraints of the problem as possible. In this case, the objective
is to find the fixed pressure and location combination of fans in
the available installation locations of the mine, in order to
provide the required airflow with minimum cost or minimum
deviations from requirements. The fitness functions will be
explained for each analysis. As the current practice in the mining
industry is to account for the maximum requirement over a
certain amount of time, the idea is to test these two different
fitness functions and demonstrate their potential to generate
savings compared to manual solutions.
The proposed representation consists of the pressure value for
each fan location. Pressures are defined as a discretised range of
values at which the fan can operate. The range for booster and
main fans are the same: ranging from 0 to 4000 pa with
discretised steps of 100 pa. Possible fan locations result from the
Spectrum Series Volume 17 265
E ACUÑA et al

MF
WF2
WF5

WF3

F2
WF4

F1

WF1

FIG 1 - Simplified schematic of the ventilation network.

physical restrictions of the mine, and are an input to the problem.

The chromosome representation, as shown in Table 1, will TABLE 1
specify a fixed pressure for each fan location that will be used by Chromosome (string).
the ventilation solver (3D-CANVENT) to generate the airflows
of the network. 1500 200 0 400
Fan fixed Fan fixed … Fan fixed Fan fixed
Population size and number of generations pressure at pressure at pressure at pressure at
position 1 position 2 position p - 1 position p
The population size, n, is the number of individuals maintained
through each generation of the GA; these individuals provide the
basis for generating improved solutions. This parameter will Equation 8). The problem with this approach is that the airflow
requirements are not included in the objective function, nor in
impact the performance of the GA, but it has to be determined
the chromosome representation, so the GA will not be able to
experimentally.
find a feasible solution. To correct this problem, a large preset
penalty term is added to the fitness when the solution is
Constraint handling infeasible. This forces the search to concentrate on the feasible
Several different constraints must be considered in the context of region.
mine ventilation, as outlined by Lowndes and Yang (2004). In
Pj ⋅ Qj
this study, the only constraint considered is to supply at least the
required airflow to the different work faces. It is understood that ∑ 1000
+ Penalty _ value (8)
j ∈FanLocations
quantity is not enough; quality is also necessary. As the intention
of this study is to evaluate the performance of the different
fitness functions, the only constraint that will be considered is where:
the airflow requirements. Quality is handled through the use of Pj is the pressure of the fan in branch j
an initial oriented network without recirculating loops and a final
Qj is the airflow in branch j
check on the directions of the flows according to the initial
network. FanLocations are the branches where the fans can be located
and Penalty_value is a large preset value.
Fitness functions
Airflow deviation fitness function
Airpower fitness function The second approach is to minimise the deviation of the airflow
A first approach to minimise the energy consumption in a mine supplied from the airflow requirement, as it is generally assumed
ventilation network is to minimise the airpower that is required that if there is less deviation from the airflow required in
to supply the airflow requirement through the main ventilation different work faces, the energy costs to supply that airflow will
systems. It is generally assumed that if there is less airpower also be reduced. In that case, the fitness function can be given as
utilised to provide the airflow required in different work faces, the sum of the square of the differences between the airflow
the amount of airflow provided will also be reduced, adjusted to requirement and the airflow provided by the fixed pressure
the demand. In that case, the fitness function can be given as the combination (see Equation 9). The problem with this approach is
sum of the airpower required by the fixed pressure fans (see the symmetry of the fitness function with respect to the airflow

266 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 AN APPLICATION OF GENETIC ALGORITHMS TO THE MAIN VENTILATION SYSTEM OPTIMISATION

requirements. As an example, if the airflow requirement is 30 m3/s population sizes ranging from 50 to 300 with a step of 50.
and the airflow supplied is 20 m3/s or 40 m3/s, then the value of Resulting values and solving times are displayed in Table 3.
the fitness function will be the same, and the GA will not be able
to differentiate between a feasible solution and an infeasible one.
To correct this problem, a large preset penalty term is assigned TABLE 3
when the solution is infeasible. This forces the search to Solving time and airpower for airpower function
concentrate on the feasible region. (hours:minutes:seconds).
AP Gen Gen

∑ (Qr j − Qs j )2 + Penalty _ value (9) Pop size 50 100 150 50 100 150
j ∈FanLocations 50 0:00:24 0:00:45 0:01:06 905.93 837.68 1002.20
100 0:00:49 0:01:16 0:01:50 919.29 767.04 805.15
where:
150 0:01:12 0:01:59 0:02:50 815.97 767.04 861.16
Q is the airflow requirement in branch j 200 0:01:43 0:02:49 0:04:05 849.24 849.24 849.24
Qs is the airflow supplied in branch 250 0:02:02 0:03:24 0:04:49 880.33 815.97 805.15
AirflowLocations are the branches where the airflow 300 0:02:30 0:04:12 0:05:41 805.15 849.24 805.15
requirements are fixed by planned activities and Penalty_value
is a large preset value.
For all 18 runs, the GA was able to find feasible solutions for
the ventilation network where the airflow requirements were
Selection, crossover and mutation operators delivered at least as required. In 15 out of the 18 runs, the GA
was able to improve the manual optimised solution, and the
The GA used is the one described in the algorithm values are presented with shaded background. The best solution
implementation section, where the selection method is roulette in terms of the airpower was 767.04 kW, and compared to the
wheel and the crossover operator is one point crossover, initial optimised solution that generated an improvement of 13 per
described by Goldberg (1989). The crossover probability is cent. The best value was found in two runs out of the 18.
60 per cent and the mutation probability is ten per cent.
Airflow deviation fitness function
Baseline For the airflow deviation fitness function approach, the same 18
runs were generated as for the airpower fitness function
Hardcastle (1983) had previously determined an initial manual
approach. Solving times and results are displayed in Table 4.
optimised solution for the ventilation network according to the
operation in the field. This allows two things: first, a check of
whether good feasible solutions are generated and second, a TABLE 4
comparison of the performance of the GA against a manual Solving time and airpower for airflow deviation function
optimised solution. (hours:minutes:seconds).
The results of the original manually optimised solution are
Dev Gen Gen
presented in Table 2. Three potential fan locations are allowed in
the mine: one main fan (MF) and two booster fan (F1 and F2). A Pop size 50 100 150 50 100 150
fourth fixed pressure booster fan represents the natural 50 0:00:23 0:00:53 0:01:19 892.49 1002.20 905.93
ventilation pressure (NVP), but it is not shown on the network 100 0:00:57 0:01:32 0:02:08 880.33 815.97 861.17
diagram. The total combined fan power determined from this
initial manual solution was 881.43 kW. 150 0:01:20 0:02:20 0:03:05 861.17 933.02 861.17
200 0:02:05 0:03:19 0:04:27 1002.20 919.29 986.75
RESULTS 250 0:02:03 0:03:50 0:05:34 933.02 933.02 861.17
300 0:02:39 0:04:12 0:05:38 861.17 971.88 767.04
Airpower fitness function
Again, for all 18 runs the GA was able to find feasible
Based on the parameters and the chromosome representation solutions for the ventilation network. For this case, in only eight
described in the previous section, the GA was implemented. runs out of the 18 runs the GA was able to improve the manual
Several runs were generated to test its ability to find feasible optimised solution, and the values are presented with shaded
improved solutions in terms of the airpower consumed, in order background. The best solution in terms of the airflow deviation
to deliver the airflow requirements to the work faces. 18 runs had a value of 0.79, and the airpower associated with that
were generated with three different generation numbers solution was 767.04 kW. The best value was found in only one
(iterations): 50, 100 and 150, and with a varying number of run out of the 18.

TABLE 2
The initial manual optimised solution.

Fan No Branch From To Fixed pressure Flow Fan

1 3 2 202 2.388 192.96 MF
2 16 72 71 2.094 56.32 F1
3 25 90 91 3.996 69.9 F2
4 6 3 2 0.131 178.52 NVP
kW 881.43

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 267
E ACUÑA et al
1050
1000
950
Airpower
900
850
800
750
0.8 0.9 1.0 1.1 1.2 1.3
Airflow deviation
FIG 2 - Airflow deviation versus airpower.
Comparison between both approaches
Values and feasibility of the solutions
Both approaches were able to generate feasible solutions for the
ventilation network with the population size and generation
number implemented. However, 83 per cent of solutions
generated with the airpower fitness function were better than the
initial manual optimised solution. The solutions generated with
the airflow deviation approach were better than the manually
optimised solution only 44 per cent of the time. Although both
approaches converged to the same ‘best’ optimised solution
(767.04 kW), the results obtained do not show correlation
between airflow deviation and airpower. Figure 2 presents the
airflow deviation versus the airpower for the airflow deviation
fitness function results. It was initially expected that a higher
airflow deviation will have higher airpower consumption, but
near the ‘best’ solution found the behaviour observed was just the
opposite except only by the ‘best’ solution found. The reason for
these results is that the fine adjustment to airflow requirements in
the work faces of the network around the ‘best’ solution found
generates more expensive solutions airpower wise, but because
the fitness function just seeks to reduce airflow deviation, it
cannot recognise whether a reduced deviation has a positive
impact on airpower at the same time.
Solving time
As expected, the solving time is similar in both approaches; one is
not clearly faster than the other. The solving time for the airpower
fitness function ranged from 24 seconds up to five minutes and
41 seconds, with a total time for the 18 runs of 43 minutes and
26 seconds. In the case of the airflow deviation fitness function
the solving time ranged from 23 seconds up to five minutes and
38 seconds, with a total time for the 18 runs of 47 minutes and
46 seconds.
CONCLUSIONS
Both the airpower and the airflow deviation fitness functions were
able to demonstrate their capacity to find improved feasible
solutions for the mine ventilation network in terms of the airpower
and the airflow required for the different work faces, and to
achieve the same ‘best’ result with an improvement of 13 per cent
compared to the baseline. In terms of solution quality, both
approaches are comparable, but the airpower fitness function
outperformed the airflow deviation fitness function in terms of
the number of times that the initial manual optimised solution
268 Spectrum Series Volume 17
was improved. In terms of the solving time, both fitness
functions are comparable as expected. Further work is required
to determine the best set of parameters for each network and
what is the impact of larger networks with a larger number of
fans in the solving time.
ACKNOWLEDGEMENTS
This research was partially supported by MITACS
ACCELERATE, Ontario, Canada. The authors would like to
thank Gary Li of Natural Resources Canada and Robert Maynard
of MIRARCO for their support integrating 3D-CANVENT with
the algorithm. The genetic algorithm was developed using the
GAlib genetic algorithm package, written by Matthew Wall at
the Massachusetts Institute of Technology.
REFERENCES
Acuña, E, Hardcastle, S, Maynard, R, Fava, L, Hall, S and Dunn, P, 2009.
The application of genetic algorithms to multiple period ventilation
systems for multilevel mine operations, in Proceedings Orebody
Modelling and Strategic Mine Planning (ed: R Dimitrakopoulos),
pp 265-270 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Bodenhofer, U, 2007. Genetic algorithms: Theory and applications
[online], Fuzzy Logic Laboratorium Linz-Hagenberg. Available
from: <http://www.flll.uni-linz.ac.at/teaching/Ga/GA-Notes.pdf>.
Calizaya, F, McPherson, M J and Mousset-Jones, P, 1987. An algorithm
for selecting the optimum combination of main and booster fans in
underground mines, in Proceedings Third US Mine Ventilation
Symposium, pp 408-417.
Goldberg, D E, 1989. Genetic Algorithms in Search, Optimization and
Machine Learning (Addison-Wesley, Don Mills: Ontario).
Hardcastle, S G, 1983. Analysis of airflow patterns and pollutant
concentrations in mine ventilation recirculation systems, in thesis
submitted to the University of Nottingham for the degree of doctor of
philosophy, volumes 1 and 2.
Hardcastle, S G and Kocsis, C K, 2007. The ventilation Challenge – A
Canadian perspective of maintaining a good working environment in
deep mines, in Challenges in Deep and High Stress Mining,
pp 519-525.
Hardcastle, S G, Kocsis, C K, Bissiri, Y and Baiden, G, 2007. Continued
development of a process simulation approach to determine airflow
requirements in the optimal design of a ventilation system, in
Challenges in Deep and High Stress Mining, pp 537-544.
Kocsis, C K, Hall, R and Hardcastle, S G, 2003. The integration of mine
simulation and ventilation simulation to develop a ‘life-cycle’ mine
ventilation system, in Application of Computers and Operation
Research in the Mineral Industries, pp 223-229 (Southern African
Institute of Mining and Metallurgy).
Lowndes, I S, Fogarty, T and Yang, Z Y, 2005. The application of
genetics algorithms to optimize the performance of a mine
ventilation network: The influence of coding method and population
size, in Soft Computing, volume 9, pp 493-506.
Lowndes, I S and Yang, Z Y, 2004. The application of GA optimization
method to the design of practical ventilation systems for multi-level
metal mine operation, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 113(1):A43-A58.
McPherson, M J, 1993. Subsurface Ventilation and Environmental
Engineering (Chapman & Hall: London).
Petrov, N N and Popov, N A, 2004. Ways of improving economy and
reliability of the mine ventilation, in Journal of Mining Science,
40(5)531-536.
Wu, X S and Topuz, E, 1998. Analysis of mine ventilation systems using
operations research methods, in International Transactions in
Operational Research, 5(4)245-254.
Xing, W and Topuz, E, 1988. Comparison of methods for determination
of booster fan locations in underground mines, in Proceedings Fourth
US Mine Ventilation Symposium, pp 355-362.
Advances in Orebody Modelling and Strategic Mine Planning I

Applying Sequence Optimisation Technology to Longwall Coal
Mining
L Rocchi1, P Carter2 and P Stone3
ABSTRACT
BHP Billiton’s Illawarra Coal (IC) operates several longwall coal
extraction systems in the Bulli and Wongawilli coal seams contained
within the Southern Coalfields of the Sydney Basin. IC has applied a
proprietary Mixed Integer Linear Programming (MILP) mine planning
optimisation tool called Blasor to rapidly evaluate the jointly optimal
underground development strategy and mining sequence for a number of
longwall mining domains. For each development scenario, the Blasor
optimiser maximises the discounted operating cash flow (DCF) as the
objective function, subject to mining, processing and transportation
capacity constraints and product blend constraints. The paper outlines how
Blasor, a planning tool developed specifically for open pit strategic
planning, has been applied to the underground longwall domain sequence
optimisation problem. It demonstrates the value of such a tool for
optimising strategic decisions through the evaluation of a set of carefully
considered scenarios, with each scenario describing a feasible underground
development and transport strategy. In this application, Blasor plays the
role of a fair valuation tool for each major scenario, in that it discovers the
most valuable extraction schedule for each development scenario, thus
establishing a sound basis for comparing the economic merits of
competing, and often structurally quite different, scenarios.
INTRODUCTION
BHP Billitons Illawarra Coal mining operations are focused on the
mining of two seams – the Bulli seam and the Wongawill seam, of
the Southern Coalfields of the Sydney Basin, centred around the
coal export facilities at Port Kembla in New South Wales,
Australia. Illawarra Coal (IC) typically produces a high quality
metallurgical coal and lesser quantities of steaming coal at annual
rates of about 8 Mt. Coal is produced from three longwall mining
systems at Appin, West Cliff and the recently commissioned
Dendrobium Colliery. Two Coal Preparation Plants (CPP) are
supplied with coal from these longwalls, one located at West Cliff,
the other located near to BlueScope Steel’s Port Kembla blast
furnaces. IC has a contract to supply 2.5 - 3.0 Mt per annum to
BlueScope, with the balance being delivered to the Port Kembla
Coal Terminal (PKCT) for transport to other Australian and
overseas customers. Resource modelling and mine planning
activities are conducted using Mincom MineScape software and
Runge XPAC software. The XPAC scheduler requires the user to
determine a ‘strategic’ path through the longwall domains, where
these comprise a number of longwall blocks. Longwall blocks are
typically 300 m wide and 1 - 3 km in length. XPAC then defines
the detailed extraction sequence which follows this strategic
direction by using a myopic search of the different legal short-term
extraction options and choosing the one which best meets the
objective and relaxed constraints. XPAC usually has a difficult
time in meeting hard blend constraints, but it can apply penalties
as a method of forcing compliance.
1. MAusIMM, Manager Resource Strategy, Illawarra Coal, BHP
Billiton Limited, Old Port Road, Wollongong NSW 2505.
Email: Luca.R.Rocchi@bhpbilliton.com
2. MAusIMM, Mining Engineering Manager, Resource and Business
Optimisation, BHP Billiton Limited, 152 - 158 St Georges Terrace,
Perth WA 6000. Email: peter.g.carter@bhpbilliton.com
3. Manager Optimisation R&D, Resource and Business Optimisation,
BHP Billiton Limited, Richard Berry Building, Melbourne
University, Melbourne Vic 3000.
Email: peter.m.stone@bhpbilliton.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
Run-of-mine (ROM) and washed coal product transportation
logistics are complex. ROM coal from Dendrobium is transported
via rail to Port Kembla, with CPP rejects transported by road for
disposal at the West Cliff Colliery. Product coal for export is then
transported by road to the PKCT. ROM coal from the Appin
Colliery is transported a short distance by road to the West Cliff
CPP, with wash plant products subsequently transported by road to
the PKCT or BlueScope Steel. Defining longwall domains within
the relevant coal seam is also a complex matter. As longwall
mining results in varying degrees of subsidence, and the
Wollongong escarpment is a heavily populated area, definition of
Coal Extraction Zones (CEZ) consistent with efficient longwall
operations and minimal public disturbance requires a considerable
mine planning effort. Typically, further data must also be gathered
about the coal resource ahead of mining, which when coupled
with the various planning approvals, can require significant
amounts of time. Underground transport options for coal produced
by a longwall are also complicated, particularly for the Appin/
West Cliff Colliery (see Figure 1) where several domains could be
transported to the surface by one of the hoisting systems, each of
which has its own cost structure and capacity constraints. The
Dendrobium Colliery is less complex, with a single coal hoisting
system (see Figure 2).
Coal qualities vary by domain within each colliery. The ranges
across all blocks in all domains are presented in Table 1. On a
block-by-block comparison, the West Cliff domain offers the
highest coking coal yields (the more valuable product), whilst
Dendrobium tends to offer the lowest. Vitrinite, ash and specific
energy values also vary quiet widely between collieries. On a
weighted average basis, the Appin domains offer the highest
coking coal yields. Tonnage splits (see Figure 3) also contribute
to the complexity of the optimisation problem – what is the
optimum mining sequence that maximises the value of these coal
resources, given product prices, the costs of mining, washing and
transporting the coal, capacity constraints of the various
components and the quality constraints for a blended product? A
problem this complex is ideally suited to the application of
Mixed Integer Linear Programming (MILP) technology on
which the Blasor software is based.
BLASOR METHODOLOGY
Stone et al (2007) have discussed the development of the Blasor
software. Essential inputs include a block model file, definition of
an operational model to describe material movements and
definition of an economic model from which discounted cash
flows are derived. In a novel application, the block model file can
be presented to Blasor as a virtual pit which excludes the
surrounding waste material otherwise required for a standard pit
optimisation. Some simplifications of the orthodox open pit
mining constraints are required particularly in relation to enforcing
mining continuity in a longwall domain once opened, otherwise
the mathematical model developed within Blasor provides a very
realistic representation of the optimisation problem and is a global
optimisation (ie objective function of maximising discounted cash
flow for the life-of-mine, subject to various constraints), as
compared to the XPAC heuristic methodology. At the core of the
Blasor optimisation engine is a commercially available mixed
integer programming solver, CPLEX, a very powerful and flexible
Spectrum Series Volume 17 271
L ROCCHI, P CARTER and P STONE

Appin 7E

Appin 7W
West Clif f 5

Appin Drif t

Appinn8
Appi 8
North Cli ff

Wes t Cliff
Appin 3 Winder & CPP

FIG 1 - Longwall domains for Appin/Westcliff collieries.

FIG 2 - Longwall domains for the Dendrobium Colliery.

272 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 APPLYING SEQUENCE OPTIMISATION TECHNOLOGY TO LONGWALL COAL MINING

TABLE 1
Coal quality variation by colliery.

Coal quality (block ranges) Units Appin Westcliff Dendrobium

Coking coal yield (min-max) % 44 - 89 70 - 87 34 - 50
Steaming coal yield (min-max) % 0.6 - 37 4 - 16 18 - 35
Vitrinite content (min-max) % 43 - 52 37 - 46 71 - 86
Ash content (min-max) % 10 - 15 10 - 14 27 - 31
Specific energy content (min-max) kcal/kg 6620 - 7030 6705 - 6990 6200 - 6490

Longwall Domain Tonnage Splits available constraint choices once an initial optimisation had been
completed. Some variables were ignored for the optimisation –
coal mined by development to establish access for the longwall
system, ventilation of the workings and the establishing of coal
37% conveying systems. It was reasoned that an optimum sequence
for the domains was the critical variable, with the selected
sequences being used to inform the XPAC data structure. This
51% specifically included mine development advance and tonnages,
thereby facilitating a more detailed calculation of cash flows.

Economic parameters and physical constraints

12%
Appin Colliery Westcliff Colliery Dendrobium Colliery
FIG 3 - Distribution of coal inventory between the collieries.
tool. Its direct application to mining optimisation problems is not
possible, however, due to the enormous number of decision
variables. To reduce the size of the problem, Blasor uses a number
of proprietary heuristic algorithms, effectively aggregating and
re-scaling space and time variables. The solution provided by
Blasor is provably optimal within a predefined tolerance.
THE OPTIMISATION PROBLEM
IC identified two elements to the overall problem. Relocation of
one longwall system in the Appin/West Cliff domains was
imminent. A second and equally important question was related
to an optimisation across all domains to determine whether the
optimisation of the Appin/West Cliff domains would be affected.
Input models
Data describing the essential physical characteristics of the coal
reserves within each longwall block was extracted from the
Mincom MineScape 4 General Mine Planning (GMP) package.
Quantity and quality data (typically the parameters in Table 1)
were then manipulated to establish a data structure suitable for
Blasor. In essence, this data structure reflected a column of cells
with the uppermost cell representing the first block to be mined
in a domain and the lowermost block, the last to be mined. This
permitted differing directions of advance within a domain to be
simulated, reflecting the differing ROM coal hoisting options.
The Appin/West Cliff Colliery complex comprised six
domains for a total of 91 blocks, averaging about 15 blocks for
each domain (range 4 - 17). The Dendrobium Colliery comprised
an additional two domains for a total of 42 blocks. Each domain
was presented to Blasor as a distinct virtual open pit, devoid of
any surrounding overburden. The mining sequence within a
domain is implicitly an open pit direction of advance – ie
downward. Constraints to force Blasor to maintain continuity
within a domain once mining has commenced are not a natural
feature of the software – this required some manipulation of the
Blasor has an extensive range of options for applying constraints
which are consistent with evaluating open pit problems. Many of
these physical constraints were not relevant to the problem at
hand. The constraints which were applied are briefly described
below:
• Mine life: the number of periods over which the optimisation
should proceed – these could be months, quarters or years.
• Discount rate: the periodic discount rate which should be
applied to the cash flows (revenues less costs) calculated to
arise from the mining and processing of each block.
• Fixed costs: a time cost specified as a time series applying to
all domains, for example, a management overhead applicable
to the aggregate of all products for a specific time period.
• Time adjustment factor: a constant specified as a time series
which can be applied to the cost of mining or processing a
block or its products to deflate or inflate the expenditure, or
adjust for foreign exchange rates.
• Total mining rate: an upper and lower value for the aggregate
mining rate in all domains in each period, for all periods
defined by the Mine Life parameter.
• Number of pits mined simultaneously: the number of open
pits (domains) which can be in operation at the same time.
• Individual pit mining rate: for each pit (ie longwall), the
upper and lower limit for the mining rate.
Products
Blasor fundamentally models mass flows and their transform-
ation from the mine to the market. Any mass flow can be defined
as a product. There are two types of products:
1. ‘real products’ are mass flows (such as washed coking coal)
that generate revenue and thus have a non-zero price
assigned to them, as well as relevant capacity and blend
constraints; and
2. ‘pseudo products’, which are zero priced arbitrarily defined
mass flows (such as ROM coal or a contaminant of washed
coal) with any combination of capacity and blend
constraints.
For any real or pseudo product, an upper and lower limit to the
mass flow in each time period for all periods defined in the Mine

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 273
L ROCCHI, P CARTER and P STONE
Life parameter can be specified. Products could also be regarded
as contaminant masses in a real product:
• Blend constraints: for any product (real or pseudo), a
specification of the upper and lower limits proportion of the
product as a constituent of the real product, described as a
time dependent variable.
• Processing rate: for each processing plant, the upper and
lower limits to the periodic rate of coal processing, for all
periods defined by the Mine Life parameter.
• Mine out constraints: for any selected pit (ie longwall
domain), a specification as to which other domains must be
completed before commencing mining in the selected
domain.
Economic parameters
• Mining costs: the total expenditure associated with the
mining of each block, calculated as the extension of the mass
of the block by the relevant unit costs per unit mass,
adjustable by a Time Adjustment Factor where so required.
• Transport costs: the total expenditure associated with
transporting mining products from the longwall to the CPP,
measured by extending the mass of the block by the relevant
unit costs per unit mass.
• Processing costs: the total expenditure associated with
processing the entire tonnage of a block in the CPP, plus the
cost of transporting the products derived from the block and
delivering them to the PKCT, adjusted by a time adjustment
factor.
• Product prices and selling costs: a time series specification
of the prices expected to be received for each product (ie
coking coal steaming coal), together with the related selling
or transaction costs. This is also specified as a time series.
Outputs
A wide variety of outputs, graphical and tabular, are possible
with Blasor. These outputs allow the user to identify when each
block in each domain is mined and where it is sent, together with
the related periodic cash flows (discounted and undiscounted)
and periodic mass flows.
Optimisation of the Appin/West Cliff domain
sequence
The need to understand the next move for the West Cliff longwall
equipment within the Appin/West Cliff domains necessitated an
initial focus on these domains. The additional complexity of two
coal hoisting solutions for two of the six domains required the
development of a matrix (see Table 2) describing the entire range
TABLE 2
Coal hoisting alternatives matrix for the Appin/Westcliff domains.
Hoisting alternative
appin3 appin7e
A wc_hoist Yes Yes
ap_hoist
B wc_hoist Yes
ap_hoist Yes
C wc_hoist
ap_hoist Yes Yes
D wc_hoist Yes
ap_hoist Yes
274 Spectrum Series Volume 17
of coal hoisting possibilities as binary decisions allowing
selection of the optimum transport path. These were not
encodeable to the Blasor mathematical model. Product blend
constraints were not applied for this evaluation on the underlying
assumption that all coal was saleable in a blend with coal from
the Dendrobium Colliery.
The optimisation was not able to be configured to reliably
enforce mining continuity within a domain once commenced,
requiring an initial appraisal and culling ahead of applying an
interactive process to specifically enforce mining continuity. In
practice, a domain could be opened for several years, then Blasor
would switch to another domain and complete mining there
before returning to complete the project in the first domain.
Results of the initial optimisation for the six domains is
presented in Table 3. Hoisting Alternative B offered a relatively
high discounted cash flow (DCF), but for a significantly
discounted ROM coal inventory compared to Hoisting
Alternatives A and D. Hoisting Alternative C also demonstrated
a low utilisation (57 per cent) of the Blasor inventory. Both of
these results needed to be considered in the context of
contractual coal sales to BlueScope Steel over a 25 year period.
Consequently, Hoisting Alternatives A and D were carried
forward with an adjustment of the individual Pit Mining Rate and
Mine Out constraints to enforce mining continuity in each
domain after mining commenced. Small changes in the key
performance indicators emerged, however Hoisting Alternative D
remained as the best solution.
Optimisation of the Appin/West Cliff/Dendrobium
domain sequence
Evaluation of alternative coal hoisting solutions was not necessary
for the Dendrobium Colliery. Consequently, the optimisation
process was able to focus on the effect of blend constraints in the
production of the Illawarra Blend. The Illawarra Blend is a
product specification for coking coal, nominating ranges for ash
content, vitrinite content and various physical characteristics such
as fluidity and reflectance. It is a natural focus point for customers,
so the sequencing of mining domains to maximise the amount of
coal mined that satisfies the Illawarra blend remains a critical
business endeavor.
As with the previous evaluation, an initial optimisation was
completed without blend constraints, applying Hoisting
Alternative D for Appin/West Cliff and without enforced mining
continuity in a domain. Mining continuity and blend constraints
were progressively introduced (see Table 4). It became clear after
the blend constraints were applied that one domain was not being
mined. A progressive relaxation of the blend constraints revealed
that the constraint for the ash content of the blended steaming coal
product was constraining the mining sequence. Enforcing mining
continuity in the absence of any product blend constraints had a
Domains
appin7w appin8 northcliff westcliff5
Yes Yes
Yes Yes
Yes Yes
Yes Yes
Yes Yes
Yes Yes
Yes Yes
Yes Yes
Advances in Orebody Modelling and Strategic Mine Planning I
 APPLYING SEQUENCE OPTIMISATION TECHNOLOGY TO LONGWALL COAL MINING

TABLE 3
Optimisation results for Appin/Westcliff domains coal hoisting alternatives.

Hoisting alternative
KPI A B C D
Cumulative cash flow Relative to Hoisting Alternative C 126% 113% 100% 130%
Cumulative discounted cash flow Relative to Hoisting Alternative C 100% 106% 100% 108%
Mine life Years 33 24 20 32
ROM coal mined % Blasor inventory 95% 72% 57% 97%
Domains mined Number 6 5 4 6

TABLE 4
Optimisation results for Appin/Westcliff/Dendrobium domains applying coal hoisting Alternative D.

KPI Case 1 Case 2 Case 3 Case 4

Mining continuity Mining continuity Mining continuity Mining continuity
not enforced enforced enforced enforced
blend constraints blend constraints blend constraints blend constraints
not applied not applied applied relaxed
Cumulative cash flow Relative to Case 1 100% 99% 87% 89%
Cumulative discounted Relative to Case 1 100% 99% 92% 94%
cash flow
Mine life Years 33 33 28 29
ROM coal mined % Blasor inventory 100% 99% 86% 88%
Proportion of ROM % Mined 100% 100% 98% 99%
washed
Domains mined Number 8 8 7 8

small impact relative to the unconstrained results. Enforcing the
product blend constraints had a very significant effect on
utilisation of the coal resource, which was partially alleviated by
relaxing the steaming coal ash content constraint. The DCF impact
of these blend constraints was of the order of several hundred
million dollars, compared to the unconstrained case.
The reasonably significant reduction in the proportion of coal
mined from the inventory presented to Blasor was also
significant. A number of blocks were mined and assigned as
waste, ie not presented to the CPP, because this returned a higher
value solution under the enforced mining continuity requirement.
This suggested that the confidence levels in coal quality data for
these blocks should be examined and appropriate actions
implemented.
CONCLUSIONS
The application of MILP technology has provided an optimised
solution (ie highest value) for two problems: both optimising
both coal hoisting, and production within blend constraints. The
value of the mining sequences derived by this methodology was
demonstrably higher compared to mining sequences derived from
the heuristic solution methodology using XPAC software.
ACKNOWLEDGEMENTS
The authors acknowledge the consent of the management of BHP
Billiton and Illawarra Coal for approval to publish this paper.
Development of Blasor technology by BHP Billiton Exploration
and Mining Technologies is also acknowledged.
REFERENCES
Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse,
P, 2007. Blended iron ore mine planning optimisation at Yandi,
Western Australia, in Orebody Modelling and Strategic Mine
Planning, second edition (ed: R Dimitrakopoulos), pp 117-120 (The
Australasian Institute of Mining and Metallurgy: Melbourne).

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 275

Combining Optimisation and Simulation to Model a Supply Chain
from Pit to Port
P Bodon1, C Fricke2, T Sandeman3 and C Stanford4
ABSTRACT
An export supply chain, beginning with the extraction of ore from a pit
and ending with the loading of this ore onto vessels at a port, is a key
component of many mining operations. These supply chains are
comprised of a number of complex subsystems such as mining, ore
processing, transportation, stockyard management and vessel loading.
Typically, the operation and performance of each of these subsystems is
analysed in isolation, with little consideration of their interaction with
upstream and downstream subsystems. In reality, stochastic and dynamic
influences that affect one of these subsystems will have flow on effects
for all other subsystems in the supply chain. Hence, evaluation of the
performance of the total integrated system needs to capture the
interaction of these subsystems. Discrete Event Simulation (DES) has
proved to be a powerful tool in modelling supply chains, capturing the
system dynamics and interactions, and evaluating the overall performance
of the integrated system.
The primary objective of mining export supply chains is typically to
maximise production capacity, ie tonnes of ore loaded onto vessels at the
port. In some mining operations, the extracted ore is blended into a
variety of products with differing characteristics before being exported.
This can be the case for ores such as coal, iron and manganese. In these
operations, an additional objective, in the form of achieving a
predetermined quality of material on the vessels, is equally important as a
measure of system performance as production capacity. The objective of
delivering a certain quality of product often conflicts directly with the
objective of maximising production capacity, resulting in an increased
level of complexity within the supply chain. In these supply chains, the
decision-making process of planning the movement and blending of ore
through the system is paramount to the overall system performance.
Capturing this complex planning process in a DES modelling language is
possible, but proves to be a very difficult and time-consuming task. Since
planning problems are often modelled and solved using an optimisation
framework, an alternative approach is to decouple the decision-making
process from the simulation model, develop a stand alone optimisation
model for it, and then integrate the two to create a holistic model of the
supply chain. This paper describes the approach taken and presents a case
study of a successful implementation on the export supply chain of PT
Kaltim Prima Coal (KPC) in Indonesia.
DISCRETE EVENT SIMULATION AND
OPTIMISATION
Discrete event simulation (DES) modelling is the process of
emulating real world operations in a controlled environment on a
computer. This DES provides a rational and quantitative process
for increasing understanding of the potential consequences of
alternate proposals. This may range from a change in operational
philosophies through to the commissioning of new infrastructure.
Hence DES modelling is a useful tool for both long-term
strategic decision-making and short-term planning and
operational decisions.
1. Senior Consultant, TSG Consulting, Level 11, 350 Collins Street,
Melbourne Vic 3000. Email: peter@tsgconsulting.com.au
2. AAusIMM, Consultant, TSG Consulting, Level 11, 350 Collins
Street, Melbourne Vic 3000. Email: chrisf@tsgconsulting.com.au
3. Consultant, TSG Consulting, Level 11, 350 Collins Street,
Melbourne Vic 3000. Email: tom@tsgconsulting.com.au
4. Manager Coal Technology, PT Kaltim Prima Coal, Sangatta, Kutai
Timur, Kalimantan Timur 75611, Indonesia.
Email: Chris.Stanford@kpc.co.id
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
Zeigler, Praehofer and Kim (2000) describe a method for
simulating a system using a discrete event system specification.
These models are constructed by considering each physical item
(train, car dumper, reclaimer, ship, etc) as a discrete entity, with
its own uniquely defined set of properties or attributes (speed,
material type, reliability, carrying capacity, etc). These entities
act out the operational activities that make up the processes being
modelled. They consume discrete periods of time for each
activity, and incur delays that can be logically induced (eg bin
empty, no rake, etc). They also use stochastic methods to
generate randomly induced delays (eg breakdown, failures, etc),
all of which are dependent on the data and operational rules that
are defined for that particular process or piece of equipment.
This combination of logical and random events is designed to
reflect the most likely operational environment. Each system
within a DES model has individual operating rules and
parameters which need to be accurately defined. In the context of
a mining operation, an export supply chain involves the
movement of ore from pit to port, via any number of subsystems.
In many mining supply chains, the operational rules regarding
the movement of ore are simply defined (eg lump material goes
to a lump stockpile, fines material goes to a fines stockpile). The
lack of product diversification in these instances means that there
are little or no blending requirements throughout the supply
chain. These simple operational rules are able to be incorporated
into DES models of the mining supply chain relatively easily,
allowing the DES model to provide a realistic representation of
the export supply chain as a whole. However, in some mining
supply chains, the process of moving ore from pit to port is
significantly more complicated. This is particularly the case
when the ore is blended into a variety of products with differing
characteristics before being exported, which can be the case for
ores such as coal, iron and manganese. For operations such as
these, day-to-day movements of ore are typically planned and
executed by groups of experienced individuals, who match
current mining stocks and stockpile levels with a shipping plan.
The decision process by which they do so is complex, and cannot
be described using a simple set of rules. This limits the ability
of a DES modelling language to precisely replicate the
decision-making process that is used in practice, and hence
provide an accurate representation of the export supply chain.
Optimisation modelling is ideally suited for analysing complex
decision-making processes, where any number of (possibly
conflicting) objectives have been identified as being desirable,
though subject to constraints such as system capacity, operational
limitations and time. One of the most powerful features of an
optimisation model is its ability to consider hundreds of
thousands of possibilities and determine the optimal decision in a
very short period of time. In the mining industry, optimisation
modelling has been widely applied in long-term mine planning,
particularly production scheduling problems and ultimate pit
design. It is also possible to apply optimisation modelling to the
problem of planning the movement and blending of ore through a
complex export supply chain, such as those described above.
This optimisation component can then be integrated into a DES
model of the entire supply chain, enabling a holistic model of the
system to be developed. There are a number of advantages to
modelling a complex export supply chain in this manner.
Automating the process of generating the plans and carrying
them out in the simulation reduces the need for human input, and
Spectrum Series Volume 17 277
P BODON et al
aids in the process of knowledge capture and retention. In
addition, a stand-alone optimisation model provides the ability to
easily modify and test alternative planning strategies in isolation.
Optimisation models also have the ability to evaluate multiple
criteria (eg product quality, demurrage, amount of ROM
rehandling), as well as explore the effect of changing priorities
on each of these objectives.
Recently, the authors developed an automated optimisation
planning engine (APE) for use in DES models of complex export
supply chains. The APE plans the movement and blending of ore
through the system and interacts with the DES model, which
attempts to enact this plan under realistic conditions, hence
providing an accurate representation of the system dynamics of
the export supply chain. The design of the APE is described
below.
DEVELOPMENT AND IMPLEMENTATION OF THE
AUTOMATED OPTIMISATION PLANNING ENGINE
Generally, a DES model of a supply chain will consider the
performance of the system over a one year time period, using a
mine plan and shipping plan for one year as inputs. The APE is
used to plan material movements on a more frequent basis, such
as fortnightly, weekly, or a number of days in advance. The time
horizon used for the planning process is an important factor in
determining the complexity of the planning problem, hence the
computation time required to solve a problem instance with the
APE. Once a short-term plan is produced, it has to be translated
into ‘tasks’ to be carried out by the simulation. The DES model
then attempts to carry out these tasks as close as possible to the
FIG 1 - Process of interaction between the simulation and optimisation models.
278 Spectrum Series Volume 17
plan, subject to real life conditions and variability. A small
amount of intelligence is required within the simulation for
dealing with unexpected occurrences such as bad weather
shutting down pits, or pieces of equipment failing. At the end of
the planning period, control is passed back to the optimisation
model with an updated set of inputs for the next planning period.
This process is then repeated. A diagrammatic representation of
the interaction between the simulation and optimisation models
is presented in Figure 1.
The aim of the APE is to plan movements of ore from pit to
ship via intermediate subsystems such as processing plants,
transportation systems (rail networks or conveying systems) and
stockyards. A feature of export supply chains in the mining
industry is the inclusion of buffers (stockpiles and queues)
between these subsystems to mitigate the impact of subsystem
performance variability on overall system performance. In the
case of a multi-pit, multi-product blended ore mining operation,
intermediate stockpiles are used for blending the ore into
products to be shipped, as well as for buffering purposes. It
follows that inputs to the APE are short-term mine and shipping
plans and the current levels in the intermediate stockpiles. The
APE will then determine the manner in which material is to be
moved through the system via the intermediate buffer stockpiles
to attempt to satisfy the shipping plan. The objective is to
maximise the throughput of material while keeping shipped
quality as close to target as possible, subject to equipment
availability constraints. It is formulated as a mixed integer linear
program involving multiple time periods. Specific complicating
concepts related to the formulation used in the APE are
discussed below.
Advances in Orebody Modelling and Strategic Mine Planning I
 COMBINING OPTIMISATION AND SIMULATION TO MODEL A SUPPLY CHAIN FROM PIT TO PORT
Non-linear constraints
A direct formulation of a planning problem involving blending
will typically feature a number of non-linear constraints. This was
the case in the original formulation proposed for use in the APE.
A period of experimentation in the use of non-linear solvers was
undertaken with little success, as solution times tended to be
unwieldy. To overcome this issue, a small number of simplifying
assumptions are made to enable the formulation to be linearised,
so that the solution time of the APE is reduced considerably
without losing significant detail in the plans produced.
Variation in stockpile quality over time
The major complicating factor in this area is the variation in
stockpile quality over time. The quality of material extracted
from a stockpile in any time period is dependent on the quality
and quantity of material added to that stockpile in the preceding
time periods. This relationship is inherently non-linear in nature.
Assuming a fixed quality in each stockpile throughout the
planning period under consideration enables this relationship to
be modelled linearly. In addition, this assumption leads to the
possibility of a further level of planning which is more tactical
than operational, ie setting the target stockpile qualities for each
time period.
Through loading
Many port operations in mining export supply chains allow for
through loading of material onto vessels. Through loading occurs
when material arrives at the port and is loaded onto a vessel
berthed at the port immediately, effectively bypassing the buffer
stockpiles in the stockyard. From a modelling perspective, this is
achieved by including a splitter at the port, which allows a
variable amount of feed to be directed to stockpiles in the yard,
and the remainder to be sent directly to the ship loading
operation. Initial modelling of the decision of the quantity of
material to send to each respective destination resulted in non-
linear constraints. This non-linearity was overcome by assuming
that through loading is always carried out when a vessel is at
berth, and that when this occurs, a fixed proportion of the feed is
split between the through loading operation and the yard
stockpiles.
Building stockpiles to completion before being
turned over
Ideally, each port stockpile is built to completion before being
turned over and reclaimed from to load vessels. This constraint is
included in the APE by applying a penalty for turning over a
stockpile before it is full. An integer variable is included in the
formulation to represent the state of each stockpile (stacking or
reclaiming), and a fixed penalty is applied when a stockpile is
turned over before it is in a reclaim state.
Solution time
Even the most sophisticated optimisation engines generally
require a significant amount of time to determine the optimal
solution to a complex mixed integer linear programming problem
with a large number of decision variables and constraints. One of
the drawbacks of the techniques used to linearise the formulation
used in the APE is that additional decision variables and
constraints are added to the formulation. The purpose of the APE
is to improve the decision-making processes within a DES model
and facilitate the generation of realistic outputs by the DES model.
Since it is used to solve many instances of the planning problem
throughout a single simulation run, it is imperative that the
solution time for the APE is not excessive. One approach explored
in an attempt to reduce solution times was to split the planning
Advances in Orebody Modelling and Strategic Mine Planning I
problem in two (planning from mine to an intermediate buffer,
then taking this solution and planning from the intermediate buffer
to the vessels). While this resulted in a small decrease in overall
solution time, the loss of the ability to effectively plan for ships
earlier in the supply chain operation resulted in a decrease in the
quality of products loaded to vessels. Therefore, this approach to
the planning problem was abandoned.
CASE STUDY – PT KALTIM PRIMA COAL
SUPPLY CHAIN
PT Kaltim Prima Coal (KPC) operates a coal mine near Sangatta
in East Kalimantan, Indonesia. Coal is mined at various grades
and blended through a series of intermediate stockpiles that are
linked by a 13 km overland conveyor (OLC) before being loaded
as multiple products onto a ship. To determine the potential
consequences of increased production and alternative production
scenarios, KPC required an understanding of the interaction
between throughput and quality, and how these are impacted by
any changes in infrastructure and/or operating policy. The nature
of planning the KPC operation to achieve contracted coal
qualities involves multiple objectives including throughput,
blending and on time delivery onto ships. An APE was
constructed that incorporated these multiple objectives, as well
as the ability to interact with the DES model. The optimisation
replicates the planning activity that is regularly performed on site
to enable the DES model to operate for an extended duration
(anywhere from one week to several years). The integration of
the optimisation within the simulation allowed KPC to link the
effects of real life uncertainties to the strategic plans being
developed. The addition and integration of the APE within the
simulation model was a complex task. In isolation, the three key
elements of this model (quantity model, quality model and
planning) are well established, however the incremental addition
of each of these elements into an integrated DES and
optimisation model exponentially increases the model
complexity. The insights gained from this complex modelling
system hold the potential to revolutionise the way the KPC
operation is planned and operated.
Overall benefit
The integrated DES model has helped KPC in making strategic
long term decisions, short-term planning decisions and also
provides the possibility of aiding the operational decisions of
creating and evaluating weekly plans.
Strategic decision-making
The primary purpose of building the integrated DES model was to
help KPC understand the likely impacts of increased production,
identify bottlenecks in the system and evaluate the effect and
feasibility of various potential future expansions of the operation.
The ability to easily change the inputs of the model enables KPC
to quickly understand the effect of upgrades to equipment, such as
increasing crusher capacity, improving conveyor rates and
reliability, and increasing ship loading capacity. In addition, the
integrated DES model provides the flexibility to test different
stockpiling configurations and examine the effect of reducing the
amount of through loading.
Short-term planning
The ability to take a short-term marketing plan in terms of
shipping demand for tonnes and quality, and evaluate the
potential of the operation to supply these tonnes given a mine
plan, has been of great benefit to KPC. The DES allows KPC to
see which plans are harder to achieve, as well as indicating the
potential bottlenecks and areas which are causing problems.
Spectrum Series Volume 17 279
P BODON et al
Operational decision-making
The optimisation component of the integrated DES model is able
to operate in stand-alone mode. This provides the ability to easily
modify and test alternative planning strategies in isolation. In
addition, further detailed modelling of the planning process that
occurs on a day-to-day or even hour-to-hour basis could be
included in this stand-alone model. This could provide sufficient
detail to enable the optimisation component to be used in the
weekly planning sessions that are held on site, where the daily
movements of ore for the next week are determined.
The Integrated DES model
The KPC Coal Chain Integrated DES model is a large, complex
model that incorporates many features to enable it to accurately
replicate the real operation. The inputs for this model include
the: key plant and equipment capability/capacity, equipment
configuration, equipment reliability (planned and unplanned
downtime), mine plan, and shipping plan (forecast). Equipment
capacity and reliability is determined through analysis of the
existing operation. Equipment configuration is based on the
current operation and may be manipulated to simulate different
operating conditions. Mine and shipping plans are supplied by
KPC in the form of mine log files and the current shipping
program spreadsheet. Scenarios are arranged by manipulating
these key input parameters to test the system under different
configurations. To enable the estimation of maximum system
capacity and to identify and quantify system bottlenecks, it is
necessary to scale the demand and supply components to ensure
that the model is tested to its capacity. The supply (mine plan) is
scaled by simply scaling the coal quantity of each block in the
mine plan, while the demand (shipping plan) is scaled by adding
ships to, or removing them from the plan. The primary aim of the
APE used within the integrated DES model is to determine a
sequence of tasks to deliver the required coal to the waiting
vessels. The DES model then follows the list of tasks developed
by the APE until such time that the tasks are completed or
replanning is required. The following is an outline of the
functionality of the APE used in the integrated model of the KPC
Coal Chain.
Planning horizon
The planning horizon used by the APE is able to be varied.
Following a testing phase it was established that a planning
horizon of 21 shifts (seven days), with plans updated every nine
shifts (three days), provided realistic results and did not require
significant solution times. This configuration enabled the APE to
provide planning based on a seven day look ahead to ensure that
both near term and longer term (seven day) objectives could be
optimised, but also meant that the DES model was unlikely to
become significantly out of synchronisation with the plan.
Optimisation formulation
The optimisation is a general linear program as described by
Winston (1987). As such, it has an objective function which in
this case is a maximisation, subject to a number of constraints, of
the general form:
Maximise ∑ x i
Subject to ∑ a i.x i <= b i
The model is formulated in Lingo using a combination of the
techniques described by Schrage (2003).
280 Spectrum Series Volume 17
Objective function
The following is a list of objectives that the APE has to optimise
against, given the constraints listed below. The importance of
each of these objectives is controlled by weighting multipliers in
the solver:
• Maximise throughput. Primarily this means to maximise the
tonnage down the OLC, however it is extended to also
maximise tonnes mined and tonnes shipped.
• Minimise deviation of both crushed ore and port stockpiles
from their assigned lower and upper bounds. Quality
deviance is calculated by creating a variable which is the
difference between the desired and actual quality, then
including this variable in the objective with a negative value
to penalise this deviance.
• Minimise deviation from each vessels’ target quality.
• Minimise deviation of each vessels’ loading time from its
arrival time.
Constraints
The following are the list of constraints that the APE must
operate within when finding a solution that optimises the
objective function.
Pits
• Only mine blocks that are available: blocks must be mined in
sequence. This is achieved by limiting the number of blocks
that are available to be planned, based on the provided mine
plan sequence. For example, blocks nominated to be mined
in the month of September cannot be mined in August unless
all of August’s blocks have been mined. Changing the bucket
size for the blocks affects how closely the model tracks the
provided mine plan. Whilst in the short-term the order that
the blocks are mined is critical, as the model moves further
into the future, so the need for strict adherence diminishes
due to uncertainties in the orebody mapping. The ore blocks
are ‘binned’ by month – finer resolution would require more
accurate mine plans extending possibly for a number of
years.
• The APE is allowed to mine blocks that are assigned to a
following time period in the input mine plan, but a penalty is
incurred. Essentially, blocks assigned to an upcoming period
would only be mined to deal with a quality issue.
• Precedence constraints between blocks are not explicitly
modelled. It is assumed that precedence requirements have
been handled in the mine plan, and excursion from the mine
plan is allowed only on a quality issue.
• Do not exceed shovel capacity in any pit and across all pits in
any shift.
• Do not mine more than the tonnage of any given block.
Crusher stockpiles
• Do not exceed crusher rates.
• Physical configuration constraints regarding which crusher
feeds to which crusher stockpile.
• Do not exceed each crushed ore stockpiles’ capacity.
• Do not exceed each crushed ore stockpiles’ maximum
reclaim rate.
• Aim to have each stockpiles’ quality within its nominated
quality range at all times.
Advances in Orebody Modelling and Strategic Mine Planning I
 COMBINING OPTIMISATION AND SIMULATION TO MODEL A SUPPLY CHAIN FROM PIT TO PORT

Overland conveyor
• Do not exceed the maximum OLC rate.
Port stockpiles
• Do not exceed each port stockpiles’ capacity.
• Aim to have each stockpiles’ quality within its nominated
quality range at all times.
• Do not exceed total port capacity.
Vessels
• Load vessels to their stated tonnage.
• Do not load vessels prior to their arrival time.
• Do not exceed the rated capacity of the ship loader.
• Aim to have each vessel within its nominated quality range.
Outputs
The following data is output from the APE to the DES model.
The DES model then uses these outputs to control (direct) the
flow of coal.
• Tonnes to crusher stockpiles: number of tonnes from each
block in the mine plan to be sent to each crusher stockpile in
each planning period. There is the potential that a block may
be split across more than one crusher stockpile.
• Crusher stockpile to port stockpile: number of tonnes to be
sent from each crusher stockpile to each port stockpile via
the OLC in each planning period. It is possible to have more
than one crusher stockpile feeding a port stockpile and more
than one port stockpile being fed from the same crusher
stockpile.
KPIs. Furthermore, testing the system at a single throughput is not
sufficient to enable the successful identification of system
bottlenecks and inefficiencies. To enable a complete under-
standing of the system performance, the combination of the
following two KPIs is the focus of the integrated DES model of
the KPC Coal Chain:
1. Quantity: tonnes moved from mine to ship and the
utilisation of the intermediate stockpiles.
2. Quality: the match between the customers’ contracted
shipments and the coal that is actually loaded onto their
vessels. Quality is measured by the gross calorific value
(GCV) of the coal.
Quantity
DES modelling allows the user to quantify the amount of extra
production from a proposed capital expansion, as well as
providing valuable insight into the auxiliary effects from any
actions. This is particularly valuable in systems that contain
many interacting components, such as the KPC Coal Chain.
Testing the sensitivity of the system performance to varying
equipment rates allows the potential benefit of changing the
operating philosophies used in this area of the supply chain to be
determined, and also provides a means of evaluating whether
each piece of equipment is, or could potentially be, a restriction
or bottleneck on the overall system. The following is an example
of the analysis that is able to be undertaken using the integrated
System Performance
Supplied Tonnes vs Demand Tonnes
Revenue and Profit vs Demand Tonnes
Shipped tonnes

• Port stockpile to vessel: number of tonnes to be sent from

US$
each port stockpile and loaded onto each vessel in each
planning period. The APE also determines which planning
period the vessels in the queue will be loaded in.
Ship plan tonnes
Example of scenario analysis Shipped Tonnes Profit Revenue

DES modelling is a complex process that has the ability to

generate a significant quantity of output results. The FIG 2 - System response to increasing demand.
interpretation of these results requires an in depth knowledge of
the model and its outputs. To best describe the performance of
the operation under varying operating scenarios requires a
statistical comparison of results and an understanding of
statistics in general. To simplify this process, a reduced number
of key performance indicators (KPIs) have been identified that
enable a simplified and more manageable understanding of the
outputs. Just as the real system has variability associated with its
performance (due to varying equipment reliability, varying times
to perform tasks and variances in the quality of ore), so too does
the DES model. For this reason, no two actual or simulated years
will ever be the same. To cope with this fact, it is necessary to
run the DES model for the same simulated period a number of
times and then calculate the mean and standard deviation of the
results for this period. Calculating the standard deviation
quantifies the effect of variability on the process, and enables the
range of results that could be expected to be produced by the
operation under similar circumstances to be gauged. System
performance is a combination of the KPIs that describe the
performance of the system in terms of its ability to load coal onto
ships in the correct quantity and quality within reasonable time
frames. Hence system performance cannot be described alone by FIG 3 - Effect of ship loader and overland conveyor rates on
any one performance indicator, but rather is a combination of shipped tonnes.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 281
P BODON et al

DES model of the KPC Coal Chain. In this case, the aim is to CPP and TBCT Inventories
examine the effect of modifying the ship loading rate, both in Increased OLC Rate
isolation and in conjunction with the option of increasing the
OLC rate.
To establish system performance sensitivity, the model is run
at a number of throughput levels, ie using shipping plans and

Tonnes
mine plans with different levels of demand for, and supply of,
tonnes of coal. This establishes a response curve for the system.
The response curve describes the system performance as the
demands on it are increased, and provides a visual quantification
of the benefit of differing system configurations at varying levels
of throughput, as shown in Figure 2. As can be seen, increasing
ship plan tonnes will increase the quantity moved through the
Time
system, however beyond a certain tonnage, it becomes harder to
meet quality targets. This causes increased penalty payments, CPP TBCT
meaning profit eventually suffers. Figure 3 shows that increasing
ship loading rate in isolation yields very little increase in the FIG 5 - Effect of increasing overland conveyor rate on stockpile
number of tonnes shipped at the lower throughput levels. This levels.
indicates that the ship loading configuration is sufficiently
capable of meeting the shipping plan at these lower throughput
levels. As the throughput is increased, it becomes clear that GCV Error - TBCT SL Berth
increasing the ship loading rate on its own yields no significant Base Case
increase in shipped tonnes. When the increased ship loading rate
is combined with an increased OLC rate, the higher throughput
levels show measurable increases over the base case
configuration, although it would appear that the majority of the
gain is provided by the increase in OLC rate.
GCV Error

Inspection of the stockpile levels (CPP = crusher stockpiles,

TBCT = port stockpiles) with the increased ship loading rate in
Figure 4 reveals that the port stockyard is running empty,
particularly towards the end of the year. The results from
increasing ship loader rates indicate that the system is unable to
supply sufficient coal to maintain adequate inventories in the port
stockpiles. This is evidenced by inspection of the scenario with Time
OLC rate increased only (Figure 5). This chart shows that the
system was far more capable of maintaining adequate port
FIG 6 - Difference between quality of coal shipped and quality of
inventory levels, and thus able to ship more tonnes and more
coal demanded over a one year time period.
easily match quality targets.

CPP and TBCT Inventories quality of coal with the loaded quality. Each dot in Figure 6
Increased SL Rate represents a ship. The y axis indicates the error in the quality of
the material that was loaded onto each ship. Therefore a positive
value indicates the model has shipped too much of a given
quality.
Continuing with the previous example of increasing the ship
Tonnes

loading rates, Figure 7 shows the shipped qualities from the

Time
CPP TBCT
FIG 4 - Effect of increasing ship loader rate on stockpile levels.
Quality
The two critical elements of the KPC Coal Chain are its ability to
deliver the correct quantity and quality of coal onto ships. In the
KPC Coal Chain operation, quality of coal is measured by gross
calorific value (GCV). The integrated DES model incorporates
the tracking of coal quality through the system and so has the
ability to measure the quality of coal loaded onto ships. To
quantify the effectiveness of the operation under various
scenarios, it is necessary to compare the contracted consignment
increased ship loading rate scenario. As a result of the low levels
of port stockpiles, it has become difficult to provide each ship
with the quality of material requested. This missmatch of quality
has a direct impact on profits, either from loss of potential
revenue when providing higher quality than required, or penalties
from customers when providing lower quality than requested. As
can be seen from this small example, there are a large number of
outputs that can be produced by a DES model, describing each
aspect of the simulated process.
CONCLUSIONS
The application of a properly developed DES model provides a
range of significant benefits in assessing an integrated export
supply chain. These benefits include the ability to assess various
operating practices, including maintenance options, through
quantification of performance. In addition to operating practices,
various capital expenditures can be compared to determine the
best infrastructure for a given system. By analysing all of these
options, the optimal capacity of the supply chain can be
determined along with the robustness of this capacity under

282 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 COMBINING OPTIMISATION AND SIMULATION TO MODEL A SUPPLY CHAIN FROM PIT TO PORT
GCV Error - TBCT SL Berth
Increased SL Rate
GCV Error
Time
FIG 7 - Effect of increasing ship loader rate on quality of
coal shipped.
uncertainty. The ability of DES to investigate outcomes over
many situations makes it ideal for risk analysis. Finally,
quantification removes the ‘gut feel’ approach and replaces it
with ‘what if’ fact based analysis.
In the case of operations that have multiple, conflicting
objectives, such as delivering a certain quality of product while
also maximising production capacity, an increased level of
Advances in Orebody Modelling and Strategic Mine Planning I
complexity is added to the export supply chain. In a mining
context, examples of ores for which this may be the case include
coal, iron and manganese. The decision-making process of
planning the movement and blending of ore through the supply
chain is paramount to the overall system performance for
operations such as these. Capturing this complex planning
process in a DES modelling language is possible, but proves to
be a very difficult and time-consuming task. Since planning
problems are often modelled and solved using an optimisation
framework, an alternative approach is to decouple the
decision-making process from the simulation model, develop a
stand-alone optimisation model for it, then integrate the two to
create a holistic model of the supply chain.
The case study of a successful implementation on the export
supply chain of PT Kaltim Prima Coal in Indonesia shows the
benefits of taking this approach to modelling for project evaluation
and strategic mine planning purposes.
REFERENCES
Schrage, Linus, 2003. Optimisation Modelling with Lingo, pp 197-320.
Winston, W L, 1987. Operations Research: Applications and Algorithms
(Duxbury Press: Boston).
Zeigler, B P, Praehofer, H and Tag Gon, K, 2000. Theory of Modelling
and Simulation, second edition, pp 176-180 (Academic Press).
Spectrum Series Volume 17 283

Direct Net Present Value Open Pit Optimisation with Probabilistic
Models
A Richmond1
ABSTRACT
Traditional implementations of open pit optimisation algorithms are
designed simply to find a set of nested open pit limits that maximise the
undiscounted financial pay-off for a series of commodity prices using a
single ‘estimated’ orebody model. Then, the maximum Net Present Value
(NPV) open pit limit is derived by considering alternate (usually only best
and worst-case) mining schedules for each open pit limit. Divorcing the
open pit limit delineation from the NPV calculation in this two-step
approach does not guarantee that an optimal NPV open pit solution will be
found. A new open pit optimisation algorithm that considers the mining
schedule is proposed. As a consequence, it can also account explicitly for
commodity price cycles and uncertainty that can be modelled by stochastic
simulation techniques. This state-of-the-art algorithm integrates Monte
Carlo-based simulation and heuristic optimisation techniques into a global
system that directly provides NPV optimal pit outlines. This new approach
to open pit optimisation is demonstrated for a large copper deposit using
multiple orebody models.
INTRODUCTION
Several open pit optimisation techniques such the Lerchs-
Grossman algorithm (Lerchs and Grossman 1965), network flow
(Johnson, 1968), pseudoflow network models (Hochbaum and
Chan, 2000) and others, involve a 3D grid of regular blocks that
is converted a priori into a pay-off matrix by considering a 3D
block model of mineral grades and economic and mining
parameters. These algorithms rely on the block pay-offs
averaging linearly, as is the case when undiscounted block
pay-offs are considered. However, the Net Present Value (NPV)
of the block pay-offs is a non-linear function of the undiscounted
block pay-offs that depends explicitly on the discount to be
applied to the individual blocks, which in turn depends on the
block mining schedule. To overcome the issue of discounting
block pay-offs, traditional implementations of open pit
optimisation algorithms are designed simply to find a set of
nested open pit limits that maximise the undiscounted financial
pay-off for a series of constant commodity prices using a single
‘estimated’ orebody model. Then, the maximum NPV open pit
limit is derived by considering alternate (usually only best and
worst-case) mining schedules for each open pit limit. This
two-step approach to finding the maximum NPV open pit limit
raises three significant issues:
1. divorcing the open pit limit delineation from the NPV
calculation does not guarantee that an optimal (maximum)
NPV open pit solution will be found;
2. NPV calculations are based on a constant commodity price
that fails to consider its time-dependant and uncertain
nature; and
3. the single ‘estimated’ orebody model is invariably
smoothed, thus it fails to consider short-scale grade
variations.
Consequently, the block model does not accurately reflect the
grade and tonnage of ore that will be extracted and processed
during mining.
1. MAusIMM, Principal Geostatistician, Golder Associates, 611
Coronation Drive, Toowong Qld 4066.
Email: arichmond@golder.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
To overcome the inadequacy of undiscounted pay-offs in
commonly used algorithms for open pit optimisation, it is
proposed to embed a scheduling heuristic within an open pit
optimisation algorithm. This may be seen as an alternative avenue
to that taken by mixed integer programming approaches (eg
Caccetta and Hill, 2003; Ramazan, 2007; Stone et al, 2007,
Menabde et al, 2007a) that may become numerically demanding
in the case of large deposits. As a consequence, uncertain and
time-dependant variables such as commodity prices can also be
incorporated stochastically into the optimisation process. This
permits strategic options for project timing and staging to be
assessed as discrete optimisation problems and compared
quantitatively and is more advanced than other recent approaches
(Monkhouse and Yates, 2007; Dimitrakopoulos and Abdel Sabour,
2007). It is also proposed to consider multiple conditional
simulations in the optimisation process such that the mining and
financial implications related to small-scale grade variations are
honoured (Menabde et al, 2007b; Ramazan and Dimitrakopoulos,
2007; Leite and Dimitrakopoulos, 2007; Godoy and
Dimitrakopoulos, 2004; Ravenscroft, 1992). By considering
discounted block pay-offs, stochastic models of commodity prices
and short-scale grade variations a more accurate discounted
pay-off matrix (revenue block model) is generated, which in turn
will yield an open pit limit that will be closer to the true optimum.
NPV CALCULATIONS WITH UNCERTAIN
VARIABLES
Calculation of the NPV for a given open pit limit relies on
estimates of numerous parameters, including (but not restricted
to) the mineral grades, extraction sequence and timing, mineral
recovery, prevailing commodity price and capital and operating
costs. All of these parameters are uncertain and should be
modelled stochastically. For example, mineral grade values by
geostatistical simulations, operating costs with growth functions
and commodity prices using long-term mean reverting models
that account for periodicity. Consequently, the cumulative
distribution of total financial pay-offs for an open pit limit can be
derived from the combination of a series of stochastic models of
mineral grades, costs, prices, recoveries, etc.
Given L potential NPV outcomes for a block (related to L
realisations of grade values, commodity prices, etc), we can
calculate the NPV for any realisation l:
B
NPVl = ∑ d l ( b j )i j (1)
j=1
and the expected NPV for L realisations:
1⎧L ⎫
NPVL = ⎨∑ NPVl ⎬ (2)
L ⎩l=1 ⎭
where:
B is the number of blocks under consideration
dl(bj) is the discounted value for block bj for the lth realisation
ij = 1 if bj falls within the open pit limit and 0 otherwise
Spectrum Series Volume 17 285
A RICHMOND
The idea being to find the open pit limit that maximises NPVL.
Additional financial goals, for example minimising downside
risk (Richmond, 2004a) could also be considered, but are outside
the scope of this paper.
ACCOUNTING FOR MULTIPLE OREBODY
MODELS
Pit optimisation algorithms found in the literature invariably
consider an orebody block model with a single grade value for
each block (or parcel). In such an approach, a simple decision
rule is used where block bj is processed using option k if
gk≤z*(bj)<gk+1,
where:
gk is the cut-off grade for processing option k (by convention
g1 = 0 and k = 1 indicates waste)
z* is the estimated grade value
To account for grade uncertainty in open pit optimisation,
Richmond (2004a) proposed incorporating L grade values for
each block. In this approach, multiple grade values zl(bj),
l=1,…,L were generated by conditional simulation and a
processing option kl(bj) was determined for each realisation.
Alternatively, conditional simulation provides short-scale grade
variations that permit local ore loss and mining dilution to be
readily accounted for in open pit optimisation by (Richmond,
2004a):
• generating geometrically irregular dig-lines (that separate ore
and waste) based on small-scale grade simulations with a
floating circle algorithm, and
• assimilating the dig-lines into large-scale geometrically
regular blocks by a novel re-blocking method.
This two-step approach accounts for short-scale grade
variation, but also provides ‘recoverable’ grade and tonnage
information for large regular blocks suitable for open pit
optimisation. In other words, the simulated grade models are
compressed without loss of accuracy so that optimisation is
computationally tractable.
AN NPV OPEN PIT OPTIMISATION ALGORITHM
For the vast majority of open pit optimisation techniques a
directed graph is superimposed onto the pay-off matrix to
identify the blocks that constitute an optimal open pit limit. To
paraphrase Dowd and Onur (1993) – each block in the grid,
represented by a vertex, is assigned a mass equal to its net
expected revenue. The vertices are connected by arcs in such a
way that the connections leading from a particular vertex to the
surface define the set of vertices (blocks) that must be removed if
that vertex (block) is to be mined. A simple 2D example is
shown in Figure 1. Blocks connected by an arc pointing away
from the vertex of a block are termed successors of that block, ie
1 2 3 4 5 in the mining schedule and receive substantially less discounting
6 7 8 9 10
11 12 13 14 15
FIG 1 - Directed graph representing 2D vertical orebody model.
286 Spectrum Series Volume 17
bi is a successor of bj if there exists an arc directed from bj to bi.
In this paper, the set of all successors of bj will be denoted as Γ j .
For example, in Figure 1, Γ 8 = {2, 3, 4}. A closure of a directed
graph, which consists of a set of blocks B, is a set of blocks
Bp⊂B such that if bj∈Bp then Γ j∈Bp. For example, in Figure 1, Bp
= {1 - 5, 7 - 9, 13} is a closure of the directed graph. The value
of a closure is the sum of the pay-offs of the vertices in the
closure. As each closure defines a possible open pit limit, the
closure with the maximum value defines the optimal open pit
limit.
For simplicity of notation, the algorithm proposed in this paper
is described for a single orebody model. The undiscounted
pay-off matrix {w(b), b∈B} typically used for open pit
optimisation is calculated as:
w(b) = tonb(vz(b)rk-ck) (3)
where:
tonb represents the tonnage of block b
v is the commodity (attribute z) value per concentration
unit
rk is the proportion of the mineral recovered using
processing option k
ck is the mining and processing cost for k ($/ton)
In practice, rk and ck commonly vary spatially and v and ck
temporally. The discounted pay-off matrix {d(b|S), b∈B},
conditional to a mining schedule S, that is required for NPV open
pit optimisation is calculated as:
d(b|S) = [tonb(vtz(b)rk-ck,t)]/(1 + DR)t (4)
where:
t is the time period in which block b is scheduled for
extraction and processing
vt ck,t are the prevailing commodity price and operating
cost at time t
DR is the discount rate
In Equation 4, discounted pay-offs are conditional to the
mining schedule as alternate schedules can be derived for the
same open pit closure. It is also important to note that, cut-off
grades and consequently the processing option k, may change in
response to commodity price and operating cost fluctuations over
time.
The traditional floating cone algorithm decomposes the full
directed graph problem into a series of independent evaluations
of individual Γ j and if the sum of the pay-offs associated with Γ j
is positive, then bj is added to Bp. However, a positive
undiscounted value for Γ j does not imply that the discounted
value for Γ j is positive. In other words, negatively-valued
successors bi of block bj that may be mined significantly earlier
may not be carried by a more heavily discounted
positively-valued bj. Furthermore, the modified schedule may
have shifted more profitable bj into later periods and additional
waste blocks into earlier periods, reducing the discounted value
of the pit. As a consequence, NPV optimisation with the FCA
must consider the directed graph problem globally rather than the
traditional independent evaluation of locally decomposed Γ j.
To allow for discounting, it is proposed that a Direct NPV
Floating Cone algorithm (DFC) proceeds as follows:
1. select the time for initial investment (start of construction) tI;
2. define a cone that satisfies the physical constraints of the
desired open pit slope angles;
Advances in Orebody Modelling and Strategic Mine Planning I
 DIRECT NET PRESENT VALUE OPEN PIT OPTIMISATION WITH PROBABILISTIC MODELS

3. define an ordered sequence of visiting blocks [1,2,…#<B] • two processing options (ore and waste), ie K = 2;
with positive w(b), by ordering the blocks bi firstly on • 60 Mt/year mill constraint;
decreasing elevation, and then for blocks with identical
elevations on decreasing value in w(bi); • 25 realisations of copper grades by Sequential Gaussian
Simulation (SGS);
4. set the open pit closure counter n = 0, the initial open pit
closure B np to a null set of blocks, and the Net Present Value • 25 stochastic simulations of future copper prices with a two
of initial open pit closure NPV n = 0; factor Pilipovic model that was modified to account for
periodicity and cap and collar aversion (Figure 2);
5. set j = 0;
• 25 stochastic simulations of operating costs with a growth
6. set j = j + 1; model (Figure 3);
7. float the cone to bj to create a new closure B np + 1 = B np + Γ j • monthly copper recoveries randomly drawn from normal
(excluding from Γ j any block that currently belongs to B np ); distribution with mean of 80 per cent and a standard
n +1 deviation of 1%2;
8. determine the schedule S for the new closure B p ;
• a fixed annual discount rate of ten per cent; and
9. calculate the discounted pay-off matrix {d(b|S), b∈B np + 1 }
using Equation 4 and the Net Present Value of the new • initial investment timings at discrete yearly intervals for five
closure using Equation 1; years.
n+1 n
10. accept the new closure if NPV - NPV > 0, whereupon
the current closure is updated into a new optimal closure, ie 350
Average
n = n + 1 and go to step 5; and 300
Simulation

11. if j<#, the number of blocks with positive pay-offs w(b), 250
then go to step 6.

Price (c/lb)
200
The deterministic floating cone algorithm presented above is
heuristic in nature and and not be optimal. Alternate Bp can be 150
generated by varying the initial investment timing (step 1), the
ordered path (step 3) and/or the mining schedule (step 8). 100

Investment timing to satisfy corporate constraints or to take 50

advantage of cyclical commodity prices can be investigated as
mutually exclusive opportunities by varying tI, which modifies 0
0 24 48 72 96 120 144 168 192 216 240 264 288 312 336 360
implicitly the mining schedule in step 8 above. For example, Time (months)
given a schedule S commencing at t = 0, the modified schedule
t‘ = t + tI . For delayed investment, the NPV for many potential
production assets will typically be reduced unless maximum FIG 2 - Thirty year future copper price simulations with mean
production/grade happens to coincide with the peak in cyclical reversion and collar and cap aversion.
commodity prices. However, for a risk averse and capital
constrained company, the shift of the capital cost into future
years may be strategically advantageous when considered in 3.5
conjunction with other mining assets. Re-initiating the test 3.3
Average
Simulation
sequence from the top of the mineral deposit each time a
3.1
positively-valued cone is found and added to the closure is
generally regarded to estimate the heuristic maximum 2.9
Waste ($/t)

undiscounted pay-off solution (Lemieux, 1979). Computational 2.7

experimentation on the ordering of blocks in step 3 above
2.5
suggested that this also holds true for the discounted case when tI
is fixed. Note that, due to re-initiation of the test sequence it is 2.3

common for B np + 1 = B np in step 7 above. For such instances, steps 2.1

8 to 10 above are ignored. 1.9
It is well known that the floating cone algorithm may not 1.7
return the maximum undiscounted pay-off solution. However, it 0 24 48 72 96 120 144 168 192 216 240 264 288 312 336 360

is used in the algorithm presented above to generate physically Time (months)

feasible solutions. The author has not investigated whether the 17

Lerchs-Grossman and network flow algorithms could be 16
Average
Simulation
substituted for the floating cone algorithm, but the non-linearity 15
of the proposed objective function may present some difficulty. 14
The computational efficiency of the proposed algorithm is 13
enhanced significantly when a simple scheduling algorithm in 12
Ore ($/t)

step 8 above is employed. However, more complex risk-based 11

scheduling algorithms to account for multiple orebody models 10

9
and production goals (eg Godoy, 2002) could be considered.
8
7
APPLICATION TO A COPPER DEPOSIT 6
5
0 24 48 72 96 120 144 168 192 216 240 264 288 312 336 360
This section demonstrates the proposed concepts for a large
Time (months)
subvertical copper deposit. The geometry and contained copper
per level are variable, but there is no strong trend. The options
considered in this study were: FIG 3 - Thirty year waste and ore processing cost simulations.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 287
A RICHMOND
Figure 2 shows 25 stochastic simulations of future copper
prices. The assumptions in this study were:
• a long-term copper price of $1.30/lb,
• the present time ($2.50/lb) was near the peak of the price
cycle,
• an average eight year copper price cycle, and
• $0.50/lb and $3.00/lb lower and upper aversion values.
Note that, as time increases uncertainty in the simulated
copper price increases and the deviation of the average simulated
value to the long-term price decreases. The average copper price
does not fluctuate symmetrically around the long-term copper
price due to the asymmetrical aversion limits. Figure 3 shows 25
stochastic simulations of waste and ore processing costs.
To assess the potential improvement in NPV against the
traditional two-stage pit optimisation approach a base case
scenario ($1.30/lb – 80 per cent recovery, $1.90/t waste cost and
$8.50/t milling cost) was run to generate a series of nested pits
using a FCA. The E-type (or average) of the 25 SGS realisations
was adopted as the single grade model as it is known to be
smoothed. The NPV for this series of pits using the base case
assumptions are shown in Figure 4 as crosses. The maximum
NPV under the base case scenario is associated with a pit closure
of 26 402 blocks. Note that, the capital cost, which could also be
modelled stochastically, was not included in this study.
5
4 FC 0 year
NPV (billion $)
3 FC 3 year
2 Base case
1 DFC 2 year
0
-1
0 20000 40000 60000 80000 100000
Pit Size (blocks)
FIG 4 - Pit size versus NPV (FCA = floating cone algorithm;
DFC = proposed direct NPV FCA).
The NPV for the FCA nested pits were also calculated using
the simulated grades, metal prices, costs and recoveries for the
six annual investment timings, shown in Figure 4. Note that:
• These curves vary substantially from the base case.
• In all instances the maximum NPV pit is significantly larger
(49 239 - 85 093 blocks) than the base case and the
maximum NPV is higher than for the base case.
• Delaying the investment from Year 3 to Year 5 results in a
higher NPV ($3.02 billion versus $2.88 billion). At first this
relationship appears counter-intuitive as costs are greater and
discounting greater. However it is related to higher Cu prices
in key production periods.
The NPV of the proposed DFC approach for the six annual
investment timings are also shown in Figure 4. Note that,
considering the mining schedule explicitly in the optimisation
process was successful in finding the maximum NPV pit in a
single run. Whilst the improvement over the maximum NPV pit
from the two-step approach that considered the stochastic inputs
was limited (usually <0.5 per cent in NPV), there was often some
difference in the pit dimension. It is likely that these differences
would be reduced further if additional pit closures had been
generated for evaluation in the two-step approach. Computat-
ionally, it was more efficient to post process a finite series of pit
288
closures than embed the scheduler in the pit optimisation process.
In the example shown, the DFC approach that generated a single
pit required around the same computational time as that required
in generating 36 nested pits by a simple FC approach.
Figure 5 shows the distribution of potential NPVs for the set
of nested FCA pits without any investment delay. As expected,
the uncertainty increases with pit size with some possibility of
negative NPVs for large pit closures. If minimising downside
financial risk is of greater importance than maximising the
NPV then the financially efficient set (frontier) of open pit
limits could be determined under a stochastic framework
(Richmond, 2004a).
9
average
simulation
8
7
6
NPV (billion $)
5
4
3
2
1
0
-1
0 20000 40000 60000 80000 100000 120000 140000
Pit Size (blocks)
FIG 5 - Pit size versus NPV distribution.
FC 1 year
FC 2 year
FC 4 year
FC 5 year
CONCLUSIONS
DFC 0 year A novel method for working with discounted pay-off matrices
DFC 1 year
during open pit optimisation was proposed. The approach used in
DFC 3 year
DFC 4 year
this study embedded a simple ore scheduler in a floating
DFC 5 year cone-based heuritic algorithm. It was a trivial exercise to further
consider multiple orebody models, local ore loss and mining
dilution, time-dependent commodity prices and costs and
120000 140000
variable metal recoveries during optimisation. As a consequence,
alternate project development timings could be strategically
assessed. Traditional evaluation of a set of nested pit shells with
constant metal prices and operating costs failed to determine the
maximum NPV pit under uncertain conditions. However, provided
that sufficient pit shells were generated and evaluated with the
same stochastic price and cost input as for the proposed algorithm
there was little difference in the maximum NPV shell derived.
Further experimentation should be undertaken to determine
whether this observation holds for more complex mining schedule
algorithms and geometrically irregular orebodies, as well as when
a smoothed block model other than the E-type of the stochastic
grade model is used to generate a series of nested closures.
This study demonstrated that uncertainty in future metal prices
and operating costs cannot be adequately captured in open pit
optimisation by simply post-processing a series of nested pit
closures with constant values. Stochastic modelling of mineral
grades, mineral recovery, commodity prices and capital and
operating costs provide an ideal platform to:
• generate an optimal pit to maximise the overall project NPV
considering geological and market uncertainty,
• determine the optimum investment and project start up
timing, and
• quantify the multiple aspects of uncertainty in a mine plan.
The example studied in this paper indicates periods of
potential financial weakness that could benefit from management
focus (eg forward selling strategies and placing the mine on care
and maintenance) prior to difficulties arising.
Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 DIRECT NET PRESENT VALUE OPEN PIT OPTIMISATION WITH PROBABILISTIC MODELS
REFERENCES
Caccetta, L and Hill, S P, 2003. An application of branch and cut to open
pit mine scheduling, Journal of Global Optimisation, 27:349-365.
Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans
under uncertainty: Can the real options make a difference?,
Resources Policy, 32:116-125.
Dowd, P A and Onur, A H, 1993. Open pit optimisation – Part 1: Optimal
open-pit design, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 102:A95-A104.
Godoy, M, 2002. The effective management of geological risk in
long-term production scheduling of open pit mines, PhD thesis,
University of Queensland, Brisbane.
Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling, SME Transactions,
316:43-50.
Hochbaum, D S and Chan, A, 2000. Performance analysis and best
implementations of old and new algorithms for the open-pit mining
problem, Operations Research, 48:894-914.
Johnson, T B, 1968. Optimum open pit mine production scheduling, PhD
thesis, University of California, Berkeley, 120 p.
Korobov, S, 1974. Method for determining ultimate open pit limits,
department of mineral engineering, Ecole Polytechnique, Montreal.
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model
for open pit mine planning: Application and risk analysis at a copper
deposit, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 116:A109-A118.
Lemieux, M, 1979. Moving cone optimising algorithm in Computer
methods for the 80s, 329-345 (SME of AIMMPE: New York).
Lerchs, H and Grossman, I F, 1965. Optimum design of open pit mines,
Bulletin of Canadian Institute of Mining, 58:47-54.
Menabde, M, Froyland, G, Stone, P and Yeates, G A, 2007. Mining
schedule optimisation for conditionally simulated orebodies, in
Orebody Modelling and Strategic Mine Planning, second edition
(ed: R Dimitrakopoulos), pp 292-310 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Menabde, M, Stone, P, Law, B and Baird, B, 2007a. Blasor – A
generalised strategic mine planning optimisation tool, 2007 SME
Annual Meeting and Exhibit.
Advances in Orebody Modelling and Strategic Mine Planning I
Monkhouse, P H L and Yeates, G, 2007. Beyond naive optimisation, in
Orebody Modelling and Strategic Mine Planning, second edition
(ed: R Dimitrakopoulos), pp 3-8 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Ramazan, S, 2007. The new fundamental tree algorithm for production
scheduling of open pit mines, European Journal of Operations
Research, 177:1153-1166.
Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
long-term production scheduling for open pit mines with a new
integer programming formulation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 310-333 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Ravenscroft, P J, 1992. Risk analysis for mine scheduling by conditional
simulation, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 101:A101-108.
Richmond, A J, 2004a. Integrating multiple simulations and mining
dilution in open pit optimisation algorithms, in Proceedings Orebody
Modelling and Strategic Mine Planning (ed: R Dimitrakopoulos),
pp 63-68 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Richmond, A J, 2004b. Financially efficient mining decisions
incorporating grade uncertainty, PhD thesis, Imperial College,
London, 208 p.
Richmond, A J and Beasley, J E, 2004. An iterative construction heuristic
for the ore selection problem, Journal of Heuristics, 10(2):153-167.
Stone, P, Froyland, G, Menabde, M, Law, B, Pasyar, R and Monkhouse,
P, 2007. Blended iron-ore mine planning optimisation at Yandi
Western Australia, in Orebody Modelling and Strategic Mine
Planning, second edition (ed: R Dimitrakopoulos), pp 117-120 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Whittle, J, 1999. A decade of open pit mine planning and optimisation –
The craft of turning algorithms into packages, in Proceedings APCOM
’99 Computer Applications in the Minerals Industries 28th
International Symposium, pp 15-24 (Colorado School of Mines:
Golden).
Spectrum Series Volume 17 289

Pushback Design of Open Pit Mines Under Geological and
Market Uncertainties
C Meagher1, S A Abdel Sabour2 and R Dimitrakopoulos3
ABSTRACT
Strategic open pit mine planning is carried out under multiple sources of
technical and financial risks. Given the uncertainty about ore grade, metal
prices and foreign exchange rates, assigning dollar values to mining
blocks at the planning time in order to optimise block destinations (mill,
stockpile or waste) is a challenge. In such an uncertain environment, there
is a probability that the actual grade, metal prices and exchange rate at the
production time will be different from those based on which an optimum
block destinations have been chosen at the planning time. In such cases,
the mine management has the flexibility to revise the original decision
regarding the block destination that has been made at the planning time
and send the block to the optimum destination based on the new
information. Another issue that affects the block value and consequently
the optimisation process is the fact that the value of a block extracted
sooner is greater than the value of the same block extracted later. This
time value of money is usually ignored in block valuation. However, it
could have a significant impact on the optimisation process.
This paper aims to quantify the value of management flexibility to
revise or alter original decisions and integrate it into the dollar value
assigned to each mining block at the planning time, while addressing the
time value of money concept. In this respect, this work outlines real
options valuation (ROV) approach for jointly handling stochastically
described geological and market uncertainties and integrating the block
destination flexibility in the process of assigning a value to mining blocks
at the planning time. The output of this ROV model is a time-dependent
discounted expected value for each block that captures the value of
management flexibility rather than a single static value. A parametric
minimum cut algorithm is then applied to produce single pushback
designs. The algorithm used intrinsically leverages the uncertainty in a set
of simulations of the orebody to produce low risk open pit pushback
designs. The method used for optimising over multiple realisations of an
orebody, as opposed to single geostatistically estimated block values and
optimising over a single orebody model, leads to an increase in the
probability that the pushbacks produced for the earlier stages of the mine
meet their expected value.
INTRODUCTION
The aim of open pit mine planning is to define optimum pit
limits and an optimum life-of-mine (LOM) production schedule
that maximise the pit value under some technical and operational
constraints. The basic input to this process is a set of block
values representing the net economic worth of each block. Based
on the estimated block values, the optimiser selects the optimum
destination of each block so as to maximise the overall pit value
under some given technical constraints. A dollar value is usually
assigned to each block by estimating the revenue of recoverable
metal at a given fixed metal price and subtracting applicable
mining, processing and other costs. Therefore, the value of each
mining block depends, among other factors, on its metal content,
metal prices and foreign exchange rates.
1. COSMO – Stochastic Mine Planning Laboratory, Department of
Mining and Materials Engineering, McGill University, FDA
Building, 3450 University Street, Montreal QC H3A 2A7, Canada.
Email: cmeagh1@cs.mcgill.ca
2. AMEC Americas Ltd, 111 Dunsmuir Street, Vancouver BC
V6B 5W3, Canada. Email: sabry.abdel-hafez@amec.com
3. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montreal QC H3A 2A7, Canada.
Email: roussos.dimitrakopoulos@mcgill.ca
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
In conventional open pit planning, block values are estimated
using a single point estimate of metal content (ore grade), a fixed
flat metal price and a fixed exchange rate (for example, Caccetta
and Giannini, 1988; Ramazan, 2007). An example of commonly
used formulas for estimating block value can be expressed as
(Whittle, 1988, 1999; Hochbaum and Chen, 2000):
V = ToGRP − ToC p − TC m (1)
where:
V = block value, $
To = tonnes of ore in the block
G = grade, unit/tonne
R = recovery
P = unit price, $/unit
Cp = processing cost, $/tonne
T = total amount of rock (ore and waste) in the block
Cm = mining cost, $/tonne
Block value estimations using current, common conventional
procedures are based on three main implied assumptions:
1. the ore grade or metal content of each block is known with
certainty,
2. market variables such as metal prices and exchange rates
are known with certainty, and
3. at the project evaluation stage it is assumed that there are
no possible future revisions of decisions related to optimum
block destinations.
This means that the decision whether to send a block to the
mill or to the dump has to be taken at the time of planning for all
blocks and no future revisions to these decisions are allowed. In
other words, at the project evaluation stage, it is assumed that all
decisions regarding block destinations have to be made upfront.
These observations have also been reported by Henry, Marcotte
and Samis (2004).
The first assumption indicates that there is perfect knowledge
about metal content of mining blocks. This clearly contradicts the
fact that some, but not complete, information about metal content
is provided by borehole data and other exploration techniques.
These data are then interpolated and/or extrapolated to develop
orebody models using geostatistical modelling techniques. This
usually results in a single, average value estimate for the metal
content of each mining block. Given the limited information and
the embedded assumption about the mineralised material between
boreholes, the probability that the estimated metal contents will be
realised is small. Accordingly, there is a high probability that the
actual metal content will be different from that used to calculate
block values. Figure 1 shows the estimated average copper grade
of a mining block at a copper deposit, as well as 20 equally
possible realisations. It is obvious that the actual block grade could
be significantly higher or lower than the estimated average. Given
the highly variant possible outcomes in Figure 1, expressing block
grade in a single number oversimplifies reality and could lead to
erroneous block valuations.
Spectrum Series Volume 17 291
C MEAGHER, S A ABDEL SABOUR and R DIMITRAKOPOULOS
1.2
1
Cu Grade, %
0.8
0.6
0.4
0.2
0
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
Simulations
FIG 1 - Possible Cu grade of a mining block.
Ignoring grade uncertainty could lead to suboptimal mine plans
with significant deviations from production targets. The ultimate
outcome has an adverse effect on the project bottom line since
project value was optimised on the basis of inaccurate inputs. The
significance of modelling and integrating geological uncertainty
into open pit mine planning has been emphasised in
Dimitrakopoulos Farrelly and Godoy (2002) and Dimitrakopoulos,
Martinez and Ramazan (2007). Uncertainty about ore grade was
identified as a critical source of risk affecting mining project
viability. Godoy and Dimitrakopoulos (2004) and Leite and
Dimitrakopoulos (2007) show through case studies that integrating
geological uncertainty in open pit mine planning significantly
reduces the risk of deviation from production targets and could
result in 26 - 28 per cent increase in project value.
The second implied assumption in conventional open pit mine
planning is that market variables such as metal prices and
exchange rates are fixed, ie do not change throughout
life-of-mine (LOM) and are known with certainty. Obviously,
this assumption is far from realistic – looking at the past history
of metal and currency markets, it is not difficult to conclude that
the probability of metal prices and exchange rates remaining
unchanged is null in both the short and the long-term. Therefore,
assuming that these variables will be constant throughout a
moderate LOM of, for instance, five to ten years will most likely
result in either over- or under-valuation of mining blocks. In both
cases, deviations from scheduled capacity and suboptimised
project value are likely outcomes. It is worth noting here that
when evaluating projects in countries other than the USA, it is
important to model the uncertainty of both metal prices and the
exchange rate. Modelling uncertainty of metal prices expressed
in the local currency and assuming that this also covers
uncertainty about exchange rate might be inappropriate. This is
because metal prices are governed by the global supply and
demand in the world markets, while the exchange rate is
governed by economic variables specific to the two countries.
The third assumption indicates that mining blocks throughout
LOM will be sent to the destinations optimised at the planning
time with no possible future revisions. This is obviously a logical
result of the first and second assumptions about certain ore grade
and market variables. In other words, if ore grade and market
variables throughout LOM were known with certainty at the
planning time, it would be possible to optimise block
destinations and consider this a final decision. In such a case, it
would be unnecessary to review the preoptimised block
destinations in the future since no changes were made to the
input data based on which these destinations were optimised.
Therefore, it could be concluded that the third assumption would
be realistic if the first and the second assumptions are realistic.
Since, as described above, block grades and market variables are
highly uncertain and consequently block values are uncertain, the
assumption that there is no revision to previously optimised
block destinations could be erroneous. To clarify this point,
292
assume for example that both the copper price at the optimisation
time and its expected long-term level equal US$2.00/lb. Figure 2
shows ten paths over a period of five years. Each path represents
a possible scenario for the future copper prices. Based on
conventional mine planning procedures, a fixed copper price of
Simulations
US$2.00 will be used to calculate block values based on which
Average optimum decisions will be made at Year 0, for each block,
whether to send it to the mill or to the dump. Consider blocks
scheduled to be extracted in Year 5 and classified as waste based
on a constant copper price of US$2.00/lb – will they be sent to
the dump in Year 5 even when the copper price follows the path
represented by Realisation-4 in Figure 2? On the other hand, for
blocks that are scheduled for the mill in Year 5 at the price of
US$2.00/lb, will the operator proceed with this schedule without
re-checking even if the actual copper price follows the path in
Realisation-6? Obviously, what happens in reality is that mine
operators usually revise previously taken decisions with time
based on the new information. This is usually carried out by
re-running the optimiser at a regular time intervals or when
major changes to the key variables take place. It is important
here to note that, standing at the early project evaluation stage,
the conventional optimiser cannot bring such a flexibility to
modify block destinations in the future.
5
Realization-4
4
Realization-7
Cu price, US$/lb
3
Realization-2
Realization-5
2
Realization-1
Expected
Realization-3
Realization-10
Realization-9
1
Realization-8
Realization-6
0
0 1 2 3 4 5 6
Time, years
FIG 2 - A sample of possible copper price realisations.
Previous work of integrating uncertainty in open pit mine
planning has been carried out by Dowd (1997), Dimitrakopoulos,
Farrelly and Godoy (2002), Godoy and Dimitrakopoulos (2004),
Leite and Dimitrakopoulos (2007) and Dimitrakopoulos,
Martinez and Ramazan (2007). Their focus was on integrating
geological uncertainty into long-term mine planning. Handling
both metal price and exchange rate uncertainty has been dealt
with in literature focusing on real options applications to mine
project valuations, which includes, among others, Brennan and
Schwartz (1985), Trigeorgis (1993), Moyen, Slade and Uppal
(1996), Kelly (1998), Moel and Tufano (2002), Monkhouse and
Yeates (2007), Abdel Sabour and Poulin (2006) and Samis et al
(2006). The focus of this real options work was on market
uncertainty – geological uncertainty was not considered. Work
that focuses on both geological and market uncertainty includes
Henry, Marcotte and Samis (2004); Dimitrakopoulos and Abdel
Sabour (2007).
This paper aims to integrate geological and market uncertainty
as well as operating flexibility to revise block destination in open
pit mine planning. This will contribute to enhancing the planning
process in a number of ways. First, stochastic models for market
variables that capture both variability and volatility will be
adopted in mining block valuation rather than using a fixed price
and exchange rate as in conventional analysis. Second, the value
of management flexibility to change decisions regarding block
Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 PUSHBACK DESIGN OF OPEN PIT MINES UNDER GEOLOGICAL AND MARKET UNCERTAINTIES

destinations in the future based on prevailing market conditions 50

is integrated and included in block value estimates. Third, pit 4

planning will be carried out using discounted block values that 40 3

account for the time value of money. The next section outlines

Block Value, 000$CAN

the proposed methodology, followed by a case study which 2
provides a practical application, and finally, conclusions and 30

recommended future extensions are presented.

20 1

UNCERTAINTY-BASED MINE PLANNING

APPROACH
10
Expected, static
The proposed approach consists of two optimisation stages: Expected, flexible

1. a dollar value is assigned to each mining block, considering

the different aspects of uncertainty and management
flexibility to revise block destination at the production time
according to the realised market variables; and
2. the second stage includes a new method for designing a
lower risk long-term mine plan based on a generalisation of
the minimum cut algorithm for ultimate pit design.
The method uses the simulation and block valuation data from
the first stage to decide when best to extract a given block to
reduce risk and increase profit.
Block valuation
As described above, ore grade and market variables are highly
uncertain and cannot be defined with a point estimate. However,
modelling uncertainty of these variables does not provide much
useful insight nor does it make significant differences in open pit
mine optimisation unless accompanied by modelling management
actions and responses to uncertainty resolutions. To illustrate the
effect of integrating uncertainty and flexibility into mining block
valuation, consider the simple case provided in Table 1 for a
FIG 3 - Comparison between the conventional and the
uncertainty-based block valuations.
value, as depicted in Figure 4. In Case 3, both ore grade and
copper price uncertainties are considered while exchange rate is
assumed to be certain. In this case, the expected block value is
112 per cent higher than that in Case 1, as indicated in Figure 3
and the range of block value is wider than that in Case 2, as shown
in Figure 4. Moving to Case 4, where uncertainty related to all
variables is considered, the expected block value is 134 per cent
higher than that in the deterministic case. It is worth noting here
that since management has the flexibility to revise block
destination according to real time information, minimum block
value is limited to the mining cost, incurred to give access to
valuable block underneath, while maximum value is unlimited and
is directly related to the three uncertain variables.
320

mining block of 10 000 tonnes scheduled for production after five Expected
270
years. For simplicity, assume that copper grade, copper price and Maximum
CAN$/US$ exchange rate can be one of the two possibilities Minimum
220
outlined in Table 1. Figure 3 shows valuation results for this block
Block value, 000$CAN

in four cases. In Case 1, no uncertainty is considered and the block

170
has been valued using conventional methods where ore grade and
market variables are assumed to be known with certainty. Block
120
value corresponding to this case is represented by the straight line
called ‘expected, static’ in Figure 3. In this case, there is only a 70
single value estimate for the block since all variables are assumed 1
2 3 4

to be known with certainty. In Case 2, only ore grade uncertainty 20

is considered while both copper price and exchange rate are
assumed to be known with certainty. Since uncertainty is -30
considered, flexibility of management to take action is integrated
as well. As shown in Figure 3, the expected block value in this
case is 62 per cent higher than that in Case 1. This result agrees FIG 4 - Valuation of a mining block in a hypothetical copper
conceptually with the findings of Godoy and Dimitrakopoulos deposit.
(2004) and Leite and Dimitrakopoulos (2007). Also, in this case,
possible block value is spread over a range rather than a single
The first step in block valuation is to model market uncertainty.
In this respect, a variety of stochastic models are used to describe
TABLE 1 the evolution of metal prices and exchange rates with time.
An illustrative example for a hypothetical block in a copper deposit. Among them, the mean-reversion model proposed by Schwartz
(1997) is widely used since it reflects the cyclic nature of metal
Mining cost, CAN$/t 2 and currency markets and estimating its parameters statistically
using historical data is straightforward. Schwartz’s model can be
Processing cost, CAN$/t ore 10
expressed as:
Process recovery 0.9
Treatment, CAN$/t concentrate 80 dS / S = η(μ − ln S )dt + σdz (2)
Refining, CAN$/lb copper 0.1
Concentrate grade, % 30 where:
Ore grade, % Cu 0.2 or 0.5 (Expected = 0.35) S is the spot value of market variable
Cu price, US$/lb 1.00 or 3.00 (Expected = 2.00) μ is the logarithm of the long-term equilibrium level
Exchange rate 0.7 or 1.25 (Expected = 0.9) η is the reversion speed

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 293
C MEAGHER, S A ABDEL SABOUR and R DIMITRAKOPOULOS

σ is the standard deviation
dz is an increment in a standard Wiener process
More details about stochastic models can be found in Schwartz
(1997) and Dixit and Pindyck (2004), among others. Based on
the stochastic model described above, a large number of correlated
realisations of market variables are generated at each period.
The second step is to evaluate each mining block, given the
generated realisations of market variables, simulated orebody
models, mining cost, processing cost and all other refining and
smelting terms. At this time, a first level optimisation process is
carried out to define the optimum destination of each block that
maximises block value. In this respect, ROV can be applied to
integrate the value of management flexibility to choose block
destinations at the actual extraction time so as to maximise block
value. A comparison is carried out between the value of sending
each block to the dump or to the processing plant. The value of a
block if it is sent to the dump is simply the mining cost:
VD = C M T (3)
where:
VD is the block value if it is sent to the dump
CM is the unit mining cost
T is the block tonnage
The block value if it is sent to the processing plant is a function
of metal(s) price(s), exchange rate, and metal content, in addition
to mining, processing, smelting and refining costs such as:
VP = F(S, MC , R, FOREX, T,C M ,C P ,C S ,C R ) (4)
where:
VP is the block value if sent to the processing plant
S is the metal price
MC is the metal content
R is the recovery
FOREX is the exchange rate
CM is the unit mining cost
CP is the unit processing cost
CS is the smelting cost
CR is the refining cost
The decision regarding optimum block destination at the
extraction time is taken so as to maximise block value:
algorithm. The ultimate pit is then broken up into continuous
smaller sections known as pushbacks, phases or cut-backs these
sections have the property that if removed in the appropriate
order they obey the engineering slope requirements for the pit.
These pushbacks are often produced using a Lerchs-Grossman
ultimate pit type algorithm with the orebody model scaled by
some parameter. Through scaling, the ultimate pit algorithm can
produce a series of nested pits which can be used as possible
choices for pushbacks.
With recent advances in simulation techniques, new methods
for producing ultimate pit limits and pushbacks are needed for
multiple realisations of the same deposit. Averaging the multiple
realisations into a single model and using the traditional
techniques does not leverage some of the upside available from
the simulations. The method described is an extension of the
network flow/minimum cut approach to ultimate pit design (for
example, Hochbaum and Chen, 2000).
Given an orebody model and economic values associated with
each block and a designation as either waste or ore, the minimum
cut algorithm for producing an ultimate pit begins by constructing
a directed graph G (for a background in graph theory and a
definition of term, see Diestel, 2005). The graph G consists of a
node for each block in the orebody model and two extra nodes, a
source node s and a sink node t. If a block bi is designated as ore
an arc from the source node s to the node representing bi is present
in G, the capacity of the arc (s, bi) is equal to the economic value
of block bi. If a block bi is designated as waste an arc (bi, t) from
the node representing block bi to the sink t is in G, the capacity of
this arc is the absolute value of the economic value of mining
block bi. Slope constraint are represented as arcs from a node vi to
a node vj if block bj must be removed prior to block bi. Slope
constraint arc have an infinite capacity.
Figure 5 shows a small 2D example of a potential block model
with economic values associated with the blocks. Figure 6 shows
the graph that would be constructed from that example, not the
slope constraint arcs have infinite capacity (and no labels).
FIG 5 - A 2D vertical cross-section of a block model.

VB = max( VD , VP S, FOREX, MC ) (5)

It is worth noting here that the block value VB is conditional

on the simulated realisation of metal price and exchange rate at
time t and the simulated orebody model. This process generates
distribution for block value at all possible extraction times. At
this stage, no operational or technical constraints such as mill
capacity and angle of slope are considered. All of these
constraints will be taken into account in the second stage of
optimisation, explained below.

Parametric minimum cut for multiple realisations

and block valuations
Traditional methods of long-term mine planning are geared
towards a single orebody model. From the single orebody model,
ultimate pit limits are calculated using the slope constraints and a
fixed economic value per block by the Lerchs-Grossman FIG 6 - The graph G constructed from the example in Figure 5.

294 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 PUSHBACK DESIGN OF OPEN PIT MINES UNDER GEOLOGICAL AND MARKET UNCERTAINTIES
A cut of a directed graph is a set of edges such that after the
removal of these edges no directed path exists between the
source node s and the sink node t. A minimum cut is the set of
edges where the sum of capacities is as small as possible over all
cuts in the graph. Many efficient algorithms are known for
computing the minimum cut of a graph in polynomial time and
they perform well in practice.
A minimum cut, F, in the graph G corresponds to a valid pit, no
slope constraint can be contained in the set F or else the sum of
capacities would be infinite but choosing the set of edges leaving
the source s is a smaller cut, a contradiction to the minimality of F.
This implies that the set of blocks corresponding to nodes not
reachable from s in G - F forms a valid pit in terms of slope
constraints. By the way the arc capacities were specific, the
minimum cut corresponds to the pit where the sum of the
capacities of arcs from the source to ore nodes left outside the pit
(s, bi ) plus the sum of capacities of arcs from the waste nodes
inside the pit to the sink (bj, t) is minimised. This minimises the
sum of ore left outside the pit plus waste inside. Since the total
sum of ore inside the whole orebody model is a constant, this is
equivalent to maximising the ore inside the pit minus the waste
inside the pit. Figure 7 shows the minimum cut in our example,
the minimum cut depicted has value 4 + (2 + 2 + 2) = 10 and
minimises the value ore left outside the pit plus the cost of the
waste in the pit.
FIG 7 - The minimum cut in the graph G.
The same algorithm can be used to produce a series of nested
pits similar to those produced by Whittle (1999). A parametric
minimum cut algorithm replaces the capacities on the arcs
leaving the source and arcs entering the sink with functions of a
FIG 8 - Minimum cut graphs with source S and sink T merged.
Advances in Orebody Modelling and Strategic Mine Planning I
single parameter λ. The capacities on arcs leaving the source s
must be non-negative non-decreasing functions of λ and the arcs
entering the sink t must be non-negative non-increasing functions
of λ. With these properties, a series of nested pits can be
produced for λ1, λ2,…, λk where λ1<λ2<…<λk efficiently. For a
small value of λ a small pit will be produced, as λ is increased
larger and larger nested pits will be created until the ultimate pit
is reached.
Given multiple orebody realisations, one would ideally like a
long-term mine plan that met production targets and achieved a
high net present value (NPV) over all realisations. A standard
approach to produce a mine plan from multiple simulations is to
average the realisations into a single orebody model and produce
the long-term mine plan in the traditional way. The problem with
this approach is that much of the information on risk due to
geologic local variability is lost when the realisations are
averaged together. For example, if you had a block worth -$3000,
-$3000 and $9000 in three different realisations then the average
value would be $1000. Similarly, a different block may be worth
$1000 in all three realisations and would be equal to $1000 when
averaged together. There is no way to distinguish that there is a
lower risk associated with the second block than the first.
Consider a small 2D example constructing the graph associated
with each simulation and merging the source nodes s and sink
nodes t, the resulting graph would look like Figure 8 (the slope
constraints have been omitted from the drawing).
Simple parametric functions on the arcs from the source will be
considered, mainly the economic value of the associated block
multiplied by λ. Since the same decision must be made in the
mine plan for a given block across all simulations, the nodes in the
different simulations should be on the same side of the minimum
cut, this can be modelled by placing bidirectional infinite arcs
between corresponding nodes. Since these bidirectional arcs have
infinite capacity they will never be in the minimum cut, so the
nodes can be merged into a single node. Arcs from s to nodes that
were merged can be replaced by a single arc with the sum of the
capacities. Arcs to t from the merged nodes can also be replaced
by a single arc with the sum of the capacities. The graph on the
left of Figure 9 depicts the merged graph from Figure 8. The graph
on the right of Figure 9 is the graph if the simulations were
averaged into a single orebody model.
The graph constructed in the fashion described contains more
information from the simulations than the graph constructed
from the averaged simulations. If one had to decide between the
block b1 that had values worth $-3000, $-3000 and $9000 in
three different realisations and the block b2 worth $1000 in all
three realisations, b2 would be chosen since the difference
between including b1 and not including it is (9000λ) - (6000),
while the difference between including b2 and not including it is
3000λ, for 0<λ<1, 9000λ-6000<3000λ. The goal of this
Spectrum Series Volume 17 295
C MEAGHER, S A ABDEL SABOUR and R DIMITRAKOPOULOS
FIG 9 - The graph on the left is constructed using the simulation
data, the graph on the right is the graph constructed if the
simulations were averaged together.
modelling is to favour blocks that have a higher probability of
being ore, this approach may be too conservative in some cases.
For instance, a block could have a greater than 50 per cent
chance of being more valuable than another but also a greater
probability of being waste, this approach will favour the block
with the higher probability of being ore. An interesting area of
further research would be to try and find a way to control the
balance between risk and reward.
By modelling the graph appropriately, parametric minimum cut
algorithms can be used to incorporate distinct prices based on the
period of extraction. Given a series of decreasing block prices
p(1,i)>p(2,i)>...>p(k,i) for an ore block i and periods 1 to k, the
graph G will have one node per period for each block i. Let v(j,i)
be the node representing block i in period j, if j < k then arc (s,
v(j,i)) will have capacity p(j,i) - p(j + 1,i) which is positive since
p(j,i)>p(j + 1,i). Let arc (s,(v(k,i)) have capacity p(k,i). If an ore
block i is removed in period j the arcs from s to v(l,i) for l<j will
cross the cut and the sum of the capacities on these arcs represents
the loss of revenue associated with waiting until period j to remove
the ore. For a waste block i and decreasing costs associated with
removing the block p(1,i)<p(2,i)<…< p(k,i) for periods 1 to k, the
arcs from v(j,i) to the sink t will have capacities p(j, i) − p(j + 1, i) .
The arc from node v(k,i) to t will have capacity p( k, i ). If the
waste block i is removed in period j, the sum of the arcs that cross
the minimum cut associated with block i equals the cost of
removing block i in period j. The graph G contains infinite
capacity arcs from every node to the corresponding node in the
subsequent period, ie (v(j,i), v(j+1,i)) for all i and j<k. This
ensures that a block that is removed in a period stays removed in
subsequent periods. Infinite slope constraint arcs exist between
nodes in the same period.
FIG 10 - Values over three different periods.
296 Spectrum Series Volume 17
Consider the following example, where we are given the
horizontal cross-section of an orebody block model’s values in
three different periods. Figure 11 shows the arcs associated with
the upper left most block. If we remove it in the first period, no
cost is added to the minimum cut. If it is removed in period 2 the
arc with value 2000 - 1800 crosses the cut and represents a loss of
a potential $200. Similarly, if we remove it in period 3 the arcs
with capacity 2000 - 1800 and 1800 - 1600 cross the cut
representing a loss of $400. If the block is left in the ground the
block all three arcs from s to the block cross the cut representing a
loss of $2000.
By converting the network described into a parametric network
by multiplying the capacity of arcs from s to a node in period i by
λi and running a parametric minimum cut algorithm choosing
appropriate values of λi for each period, one can produce a series
of nested pits over the multiple periods and choose the pits that
best meet the production schedule requirements as the pushback
design. The choice of λi's can be accomplished by a Lagrangian
relaxation approach. Where the λi's for which the pushbacks
closely match the desired requirements are chosen. The two
generalisations of the minimum cut algorithm, to multiple
realisations of the orebody model and to block evaluations based
on time of extraction, can combined to produce a set of pushbacks
with a lower risk than the traditional single orebody model
approach.
CASE STUDY – APPLICATION TO A
COPPER DEPOSIT
This section provides an application of the proposed procedures
explained above to a copper deposit. The deposit is located in a
typical archean greenstone belt. The region consists predominantly
of mafic lavas with lesser amounts of intermediate to felsics
volcaniclastics. Rocks are moderately deformed with a prominent
cleavage subparallel to what is considered to be the original
bedding, an E-W trend with average 64° south. The deposit itself
is in a sequence of moderately to strongly foliated, sulfidic, mafic
to intermediate volcanic rocks, which have been intruded by
numerous subvolcanic felsite and feldspar porphyry and/or
intermediate volcanic tuff, with size ranging from lapilli to
agglomerate, within a strongly chloritic and biotitic matrix. It can
be traced over a strike length of 1.5 km with a thickness varying
from a few metres to more than 75 m. Mineralisation consists of
about ten per cent sulfides, mostly chalcopiryte, pyrite and
pyrrhotite, occurring as disseminations, streaks and stringers
apparently controlled by the strong rock cleavage. The geological
database is compounded by 185 drill holes with ten metre copper
composites in a pseudo-regular grid of 50 m × 50 m covering an
approximately rectangular area of 1600 × 900 m2 – the average
dip is 60° north. Using the geological information, one
mineralisation domain is defined and modelled through a
geostatistical study. In this example, there are two sources of
market uncertainty, related to copper price and US$/CAN$ rate.
This market uncertainty is modelled by generating 20 000
Advances in Orebody Modelling and Strategic Mine Planning I
 PUSHBACK DESIGN OF OPEN PIT MINES UNDER GEOLOGICAL AND MARKET UNCERTAINTIES

FIG 11 - Arcs associated with the upper left most block.

correlated realisations of copper prices and exchange rates using CONCLUSIONS

the mean-reversion model in Equation 2. Uncertainty about
orebody is modelled by simulating 20 orebody models using This paper outlined a simulation-based procedure for integrating
conditional simulations. Block valuations are carried out using the both geological and market uncertainties into open pit mine
procedures explained above and the data in Table 2. Figure 12 planning. This two-level optimisation procedure combines a
shows discounted values for a sample of mining blocks using both number of uncertainty modelling, advanced financial valuation
and open pit planning algorithms. As a first level optimisation,
the conventional and the proposed uncertainty-based procedure
management flexibility to change or revise block destination was
provided that these blocks will be extracted in year 10. It is integrated into block values. The described pushback design
obvious, in this case study, that conventional block valuation algorithm can use the geological and market uncertainty to
method tends to undervalue mining blocks since it ignores produce lower risk long-term mine plans. Application to a case
flexibility to revise block destination. As shown in Figure 12, the study of a Canadian copper deposit showed significant differences
difference between the two sets of valuation ranges, in block value estimates between the conventional and the
approximately, from ten per cent to 50 per cent and it is inversely proposed procedures.
related to block value. However, it is worth noting here that in
some cases, conventional procedure might overvalue mining ACKNOWLEDGEMENTS
blocks.
The work in this paper was funded by NSERC CDR Grant
TABLE 2 335696 and BHP Billiton, and the members of the COSMO
Stochastic Mine Planning Lab – AngloGold Ashanti, Barrick,
Data of the illustrative copper mine.
BHP Billiton, De Beers, Newmont, Vale and Vale Inco.
Number of blocks 21 3001
Block tonnage, t 10 800 REFERENCES
Mining cost, CAN$/t 2 Abdel Sabour, S A and Poulin, R, 2006. Valuing real capital investments
using the least-squares Monte Carlo method, The Engineering
Processing cost, CAN$/t ore 10
Economist, 51:141-160.
Recovery 0.9 Brennan, M J and Schwartz, E S, 1985. Evaluating natural resource
Treatment, CAN$/t concentrate 80 investments, Journal of Business, 58:135-157.
Caccetta, L and Giannini, L M, 1988. An application of discrete
Refining, CAN$/lb copper 0.1
mathematics in the design of an open pit mine, Discrete Applied
Concentrate grade, % 30 Mathematics, 21:1-19.
Discount rate, % 8 Diestel, R, 2005. Graph Theory – Graduate Texts in Mathematics, third
edition, Volume 173 (Springer-Verlag: New York).
Long-term copper price, US$/lb 1.60
Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans
Long-term exchange rate 0.9 under uncertainty – Can the real options make a difference?
Resources Policy, 32:116-125.
Dimitrakopoulos, R, Farrelly, C T and Godoy, M, 2002. Moving forward
160 60 from traditional optimisation – Grade uncertainty and risk effects in
140
open pit design, Transactions of the Institutions of Mining and
50
Conventional block valuations Metallurgy, Mining Technology, 111:A82-A88.
Block Value, 000$

120 Real options valuation

Difference, %
Dimitrakopoulos, R, Martinez, L S and Ramazan, S, 2007. A maximum
Difference, %

40
100
upside/minimum downside approach to the traditional optimisation
80 30
of open pit mine design, Journal of Mining Science, 43:73-82.
60
20 Dixit, A K and Pindyck, R S, 1994. Investment Under Uuncertainty
40 (Princeton University Press: New Jersey).
10
20 Dowd, P A, 1997. Risk in minerals projects – Analysis, perception and
0 management, Transactions of the Institutions of Mining and
1 2 3 4 5 6 7 8 9 10
Metallurgy, Mining Technology, 106:A9-A18.
Mining Blocks
Godoy, M C and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling, SME Transactions,
FIG 12 - Valuation results for a sample of ten blocks. 316:43-50.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 297
C MEAGHER, S A ABDEL SABOUR and R DIMITRAKOPOULOS
Henry, E, Marcotte, D and Samis, M, 2004. Valuing a mine as a portfolio
of European call options – The effect of geological uncertainty and
implications for strategic planning, in Geostatistics (eds: O
Leuangthong and C V Deutsch), pp 501-510 (Banff: Canada).
Hochbaum, D S and Chen, A, 2000. Performance analysis and best
implementations of old and new algorithms for the open-pit mining
problem, Operations Research, 48:894-914.
Kelly, S, 1998. A binomial lattice approach for valuing a mining property
IPO, Quarterly Review of Economic Finance, 38:693-709.
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model
for open pit mine planning – Application and risk analysis at a
copper deposit, Transactions of the Institutions of Mining and
Metallurgy, Mining Technology, 116:A109-118.
Moel, A and Tufano, P, 2002. When are real options exercised? An
empirical study of mine closings, Review of Financial Studies,
15:35-64.
Monkhouse, P H L and Yeates, G, 2007. Beyond naive optimisation, in
Orebody Modelling and Strategic Mine Planning, second edition (ed:
R Dimitrakopoulos), pp 3-8 (The Australasian Institute of Mining
and Metalurgy: Melbourne).
Moyen, N, Slade, M and Uppal, R, 1996. Valuing risk and flexibility – A
comparison of methods, Resources Policy, 22:63-74.
298 Spectrum Series Volume 17
Ramazan, S, 2007. The new fundamental tree algorithm for production
scheduling of open pit mines, European Journal of Operations
Research, 177:1153-1166.
Samis, M, Davis, G A, Laughton, D and Poulin, R, 2006. Valuing
uncertain asset cash flows when there are no options – A real options
approach, Resources Policy, 30:285-298.
Schwartz, E S, 1997. The stochastic behaviour of commodity prices –
Implications for valuation and hedging, Journal of Finance,
52:923-973.
Trigeorgis, L, 1993. The nature of option interactions and the valuation of
investments with multiple real options, Journal of Financial and
Quantitative Analysis, 28:1-20.
Whittle, J, 1988. Beyond optimisation in open pit design, in Proceedings
Canadian Conference on Computer Applications in the Mineral
Industries, pp 331-337 (Balkema: Rotterdam).
Whittle, J, 1999. A decade of open pit mine planning and optimisation –
The craft of turning algorithms into packages, in Proceedings
APCOM ’99 Computer Applications in the Minerals Industries
28th International Symposium, pp 15-24 (Colorado School of
Mines: Golden).
Advances in Orebody Modelling and Strategic Mine Planning I

Faulted Geological Model Simulation of the Resolution Porphyry
Copper Deposit
G Verly1, K Brisebois2, W Hart3 and J Hammitt4
ABSTRACT
A significant portion of the risk in a mining operation is tied to the
geologic model that is being used to constrain the resource and reserve
estimates. Yet little has been published on geological model simulation in
the mining industry, perhaps because the original geologic setting has often
been disturbed by metasomatism and tectonism, making it often difficult to
interpret and simulate. One general method to realistically simulate the
geology consists of first establishing a priority ranking of the geological
features and then simulating them one at a time. This can be done by
starting with the interpreted outlines or shapes and incorporating the
uncertainties as specified by the geologist during the simulation process.
Two steps that could be part of this methodology are considered in this
paper: simulation of fault surfaces, and simulation of layered rock types
within the simulated faulted blocks. The algorithm consists of a series of
sequential Gaussian simulations that are merged together after various
rescaling to account for non-stationary uncertainties. A detailed description
of the procedure is given. Examples of results obtained on a porphyry-style
copper deposit are provided.
INTRODUCTION
A significant portion of the risk in a mining operation or study is
tied to the geological model that is being used to provide domains
for mineral resource and reserve estimation and/or conditional
simulation. Simulating the geology is a common procedure in the
petroleum industry but is rather the exception in the mining
industry. Perhaps one reason is that the original geological setting
of a mineral deposit has often been disturbed by metasomatism
and tectonism, making realistic simulations of the geology
difficult to produce. Some of the methods that can be used to
simulate different aspects of the geology are boolean modelling
(de Fouquet et al, 1989), indicator simulation (Alabert, 1987),
plurigaussian simulation (Armstrong et al, 2003), probability field
with local means (Srivastava, 2005) and potential field (Chiles et
al, 2007).
The case considered in this paper is the Resolution
porphyry-style Cu-Mo deposit, located in Arizona, USA. The
deposit is deep and sparsely drilled. The geological interpretation
is complex and contains faults, metamorphosed rocks of
sedimentary origin, intrusions, breccias and alteration assembl-
ages. One general method to realistically simulate such complex
geology consists of establishing a priority ranking for each of the
geological features and simulating them one at a time, starting
with the interpreted outlines or shapes and incorporating the
uncertainties as specified by the geologist (Verly, Bridebois and
Hart, 2008). Details on how to simulate a fault block model
together with faulted rock types are presented in this paper along
with some results.
1. AMEC, Suite 400, 111 Dunsmuir Street, Vancouver BC V6B 5W3,
Canada. Email: georges.verly@amec.com
2. AMEC, 780 Vista Boulevard, Sparks NV 89434, USA.
Email: ken.brisebois@amec.com
3. Resolution Copper Mining, LLC 102 Magma Heights, Superior AZ
85273, USA. Email: william.hart@riotinto.com
4. MAusIMM, Geologist, Kennecott Exploration Company, Rio Tinto,
224 North 2200 West, Salt Lake City UT 84116, USA.
Email: jay.hammitt@riotinto.com
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
GEOLOGY
Deposit geology
The Resolution Cu-Mo deposit is late-Cretaceous to early-
Tertiary in age and is hosted within a buried, fault-bounded
sequence of Paleozoic and Precambrian sedimentary strata,
Precambrian diabase sills and Cretaceous-aged layered
volcaniclastic and siliciclastic rocks. Host strata are faulted and
have been intruded by porphyry bodies of late Cretaceous age.
Mineralised breccia bodies have also been identified, with some
being spatially related to faults and porphyry intrusions (Hammitt
and Ballantyne, 2007). The mineralised rocks are buried
unconformably beneath a 1000 m to 1500 m thick sequence of
barren sediments and volcanic rocks of Tertiary age. Figure 1
shows a plan view and two vertical sections of a portion of the
2007 geological interpretation.
Fault interpretation and uncertainty assessment
Figure 2 illustrates a common method used by geologists for the
identification of simple block bounding faults where folding is
assumed to be minimal, using a distinctive lithologic ‘marker’
intersected in multiple drill holes (eg a formation contact).
Stage 1
Distinctive lithologic marker intercepts in each drill hole are
visually identified in section view (red symbols) and grouped
based on their apparent relative degree of colinearity (or
coplanarity).
Stage 2
A 2D marker horizon is interpreted and constructed, defined by a
smooth line (red dashes) connecting all controlling data points for
the marker within each geometric grouping. Zones of abrupt and
significant changes in position (eg elevation) between different
segments are suspected locations for fault displacement. Fault
planes are then interpreted and constructed within these suspect
zones based on numerous geological principles including:
• observed apparent sense of displacement of the marker
horizon;
• actual fault intercepts in one or more drill holes (‘true snap
points’);
• typical fault geometries observed nearby on surface or in
underground workings; and
• fault geometries predicted by well-established structural and
tectonic field studies within a similar geologic terrane, or
predicted by laboratory experiments on rock.
Stage 3
The spatial uncertainties of faults are assessed purely by distance
from controlling data points. Multiple fault snap points that are
spatially aligned may greatly reduce the uncertainty of a given
fault interpretation. Simple surfaces are then constructed that
define the limits of all permissible fault positions (blue dashed
lines).
Spectrum Series Volume 179 299
G VERLY et al

E=206
N=1006

N=1006

E=206

High Cu Shell

FIG 1 - Plan view and vertical sections through the geological interpretation. The black dots on the plan view indicate the drill hole intercepts.

FIG 2 - Diagram illustrating the methodology for the interpretation of simple faults (Stage 2, solid blue lines) in section and plan view, using
only lithologic markers intersected in irregularly-spaced or widely-spaced drill holes (Stages 1 and 2). Stage 3 illustrates the potential
degree of spatial uncertainty (orange polygons) inherent in the resulting fault model where all possible fault positions and orientations have
equal probability (dashed blue lines).

300 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 FAULTED GEOLOGICAL MODEL SIMULATION OF THE RESOLUTION PORPHYRY COPPER DEPOSIT
Faulted layered rock type interpretation and
uncertainty assessment
As described in the previous section on fault interpretation, the
boundaries for the Faulted Rock Types (layered rocks) are
interpreted and constructed on sections or in three dimensions by
first identifying the controlling drill hole data points for each
major stratigraphic horizon, and then constructing smooth lines,
polygons, or surfaces that pass through these controlling points
and enclose lithologically distinct domains. As with the faults,
every attempt is made to honor established geological principles of
stratigraphy. The resultant interpreted domains for the Faulted
Rock Types are then modelled as triangulations. Based on existing
information from both drilling and surface mapping, analysis of
the geological interpretation is performed far from drill hole
intercepts – the limit of variation for the vertical position of
domain boundaries was assessed to be ±30 m. For thickness, the
limit of variation was assessed as ±20 per cent of the interpreted
thickness.
FAULT MODEL SIMULATION
Seven faults defining eleven fault blocks have to be simulated. The
general procedure consists in simulating one fault at a time as
a 2D surface and flagging a 3D grid accordingly. The seven
3D grids corresponding to the seven faults are then merged
together using the proper priorities. Each 2D surface simulation is
constrained within a variable and asymmetrical uncertainty band-
width assessed by the geologist. The simulation methodology is
further described below and is illustrated in Figures 3 and 4.
For each of the 7 faults
Step 1: Pre-simulation processing
1a) Get original fault information:
- Fault surface (triangulation)
Drape the fault surface (dense grid)
- Uncertainty information (series of points with 2 max. distances from
interpretation)
- Snap points
1b) Get fault plane + offsets:
- Work out approximate fault plane
- Work out fault surface offsets from fault plane (draped grid)
1c) Rotate the fault information:
- Work out rotation to bring the fault plane to horizontal
- Rotate the offset information
- Rotate the uncertainty information
- Rotate the snap points
1d) Grid the rotated fault information:
- 2D horizontal grids
- Grid the offsets and the uncertainties
Step 2: 2D simulation
2a) 2D stationary simulation, conditional to snap points
2b) Rescale the simulation to account for variable uncertainty
2c) Add the offsets 100 rotated fault surfaces
Step 3: 3D simulation
3a) Rotate the 3D grid to be simulated
3b) Flag the rotated grid nodes
- Flag as above/below the simulated surfaces
100 fault simulations coded as 0/1 3D grids
Merge the simulated 3D grids
Merge grids with proper priorities
100 simulated fault grids, 7 faults / 11 fault blocks per realisation
FIG 3 - Faulted model simulation process.
Presimulation work
The available information for one fault consists of the fault
surface, the uncertainty information and three snap points for one
fault. The fault surface is modelled as a triangulation. The
uncertainty information is provided as a series of locations on the
fault surface with two maximum possible envelopes, one on each
side of the fault. The spacing between the locations is about 500 m
× 500 m. In a first step, the fault triangulation surface is sampled
on a dense grid using a draping technique. The approximate plane
of the fault is identified, and the normal offsets of the fault surface
Advances in Orebody Modelling and Strategic Mine Planning I
to the plane are calculated. These kinds of manipulations are
readily available in several resource modelling packages. Note that
the offset terminology is used for the distance from the interpreted
fault surface to the fault plane. It is not the displacement of the
fault as in normal geological parlance.
The fault plane is rotated to the horizontal. The coordinates of
the fault information (surface offsets from the plane, uncertainty
and snap points) are similarly rotated. The reason for the rotation
is that most 2D simulation algorithms do not handle 2D planes of
any orientation. Furthermore, the post-simulation grid manip-
ulations are simplified in the rotated space. After rotation, the
fault surface offsets are vertical distances from the horizontal
plane. Because the angle of the fault surface with the fault plane
at any location is always small, the maximum possible offsets at
the recorded uncertainty locations are assumed to be vertical
after rotation. The fault surface is clipped by the other faults.
Different realisations will result in different clipping. To avoid
gaps between the simulated fault blocks, the simulated 2D grid
needs to be extended sufficiently beyond the extent of the rotated
interpreted fault surface. The rotated fault surface offsets and the
vertical uncertainties are estimated on a dense 2D grid (Step 1d,
Figure 3). For the offsets, the estimation method does not matter
much as the ‘draped’ information is very dense. This
information, however, was available only on the clipped fault
surface and was estimated within the rotated outline of the
surface. The estimated values had to be extended to the limit of
the 2D horizontal grid. For the uncertainty, the available points
are much farther apart and ordinary kriging was used with a
spherical model with no nugget and very long ranges. Inverse
distance could have been used as well. No extensions were
necessary because the information was available for unclipped
surfaces. Figure 5 shows the estimated uncertainties on one side
of one fault.
Simulation
The rotated fault simulation is achieved in three steps:
1. conditional simulation of standard normal scores,
2. rescaling the simulated normal scores to account for the
non-stationary and asymmetric uncertainty bandwidth, and
3. adding the offsets to the rescaled normal scores.
The conditional simulation is sequential Gaussian. A unit-sill
Gaussian variogram model with no nugget effect and a 1000 m
isotropic range has been used to ensure a very smooth surface. The
1000 m range was chosen after some experimentation. At this
stage, 100 realisations of ‘standardised’ fault surface deviations are
available. Theoretically, the 100 simulated values at a given
location are standard normally distributed except within the zone
of influence of a snap point. The simulated values of one
realisation should also be standard normally distributed if the
surface extent is very large with respect to the variogram range.
Figure 6 shows the validation statistics for one fault surface. The
top left histogram (a) shows that the mean simulated value is very
close to zero as expected, whereas the standard deviation of a
given realisation is on average 0.66, which is lower than 1. This is
due to three snapping points that significantly reduce the possible
fluctuations and also due to the extent of the simulation being not
very large with respect to the 1000 m variogram range. The
bottom left graph (c) indicates that the variogram is reproduced by
the simulation even though significant fluctuations are observed
from one realisation to the next (d). Alfaro (2008) notes that wild
variogram fluctuations between realisations are to be expected
because the Gaussian variogram is not microergodic. The
probability plots (b) of the realisations are not straight lines
because the variances are less than one.
A local rescaling of the simulated normal scores is needed to
account for the non-stationary uncertainty. If the uncertainty was
Spectrum Series Volume 179 301
G VERLY et al

1) Rotate
a) Fault 4) Flag rotated 3D grid
F1 2) Create dense 2D grid

3) Simulate fault in 2D
a) simulate deviations
b) rescale deviations to account
Fault for the uncertainty
Plane c) add fault offsets
F1

Horizontal

b) 3D grid

5) Rotate back the grid

F1 F2 F1 F2
6) Merge
grids
with
proper
+ =
priorities

FIG 4 - Faulted model simulation main steps.

Rotated Fault -- Fault SE Side Uncertainty

(grey lines). Note that the offsets from the horizontal have not yet
400 600 800 1000 1200 1400 1600 1800 2000 2200 2400 2600 been added, which is why the interpreted fault trace lies at
deviation = 0. The final simulation step is adding the fault offsets
1000
1000

40 14 40 49 74 202
onto the rescaled fault fluctuations, resulting in 100 rotated
800
800

simulated fault surfaces on a dense 2D grid.

600
600

Post-simulation work
Down-Dip
Down-Dip

400
400

34 4 6 14 30 17 77 206
Two sets of coordinates of the simulated 3D grid nodes are
available:
200
200

1. original unrotated, and

0
0

51 37 46 40 104 281 2. rotated as per the fault rotation.

-200
-200

400 600 800 1000 1200 1400

Along Strike
1600
Uncertainty (m)
3D Grid / Fault Intersept
0. 250.
FIG 5 - Gridded uncertainty on one side of the fault after rotation.
symmetrical on each side of the fault, the rescaling factor at a
given location would be a third of the maximum distance
estimated for the uncertainty at that location – meaning that the
standard deviation of the rescaled simulated normal scores is a
third of the maximum distance. This ensures that approximately
99.75 per cent of the simulated fluctuations are within ± the
specified maximum (± 3 standard deviations). Simulated
fluctuations outside that interval are reset to the maximum
distance. The uncertainty at a given location, however, is
asymmetrical, and two rescaling factors are used, depending on
the sign of the simulated normal scores. The distribution of the
rescaled simulated normal scores consists of the halves of two
normal distributions with same zero mean but different standard
deviations. The two standard deviations are one third of the
maximum possible distances at that location. Figure 7 shows an
example of such a distribution with 10 m and 15 m ‘standard
deviations’, corresponding to maximum fluctuations of -30 m and
+45 m, respectively. Note that the mean of the resulting
distribution is not zero anymore.
One hundred realisations of fluctuations around the fault
interpretation are available at this stage. Figure 8 shows one
section with the trace of one rotated fault interpretation (red line
at deviation = 0), the maximum possible deviations (blue lines on
top and bottom) and 20 realisations of the simulated fluctuations
1800 2000 2200 2400 2600 The second set of coordinates is used to flag the rotated grid as
"Samples" "Snap Points"
above or below the simulated rotated fault, which is equivalent to
flagging the original grid as one or the other side of the un- rotated
fault. The previous steps – presimulation work, simulation, 3D
grid flagging – are repeated for each fault. When this is
completed, seven simulated 0/1 fault indicator grids are available,
with 100 realisations each. The final simulated fault block model
is obtained by merging the seven individual fault simulated models
together with the proper priorities. Figure 9 shows eight realis-
ations of the simulated fault model at mid-elevation in the model.
Results and discussion
Table 1 shows the statistics for the simulated fault block models.
To preserve confidentiality of the original data, each of the
interpreted fault block volumes has been reset to 100. The table
shows that the interpreted (modelled) fault block volumes are not
always reproduced. Differences from -14 per cent to +10 per cent
are observed. These differences are due to the asymmetric fault
uncertainties. Some simulated fault blocks have a greater chance
to become smaller than interpreted and vice versa. The table also
shows there can be very large differences between two realisations
for a given fault block. For example the average simulated volume
for Fault Block No 1 is 89, with minimum and maximum values
of 20 and 135. The coefficient of variation (CV) is 28.3 per cent.
Assuming a normal distribution, there is a 95 per cent probability
that the simulated Fault Block No 1 volumes are within ± 56.6 per
cent of the average.
The approach used to simulate the faults is considered
reasonable. Indeed, the results are statistically consistent with the

302 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 FAULTED GEOLOGICAL MODEL SIMULATION OF THE RESOLUTION PORPHYRY COPPER DEPOSIT

Fault Simulation - Simulated NSCO Checks

0.300 99.99
A) B)
99.9
99.8
0.250
99
98
95

Cumulative Probability
0.200 90
80
100 simulations 70
60
0.150 50
Nb. of data 5,109,700 40
mean 0.060 30
20
y
c
n
u
q
re
F

stdv 0.663
0.100 10
max 4.168 5
min -3.137
2
0.050 1
100 Simul.
0.2 Mean Simul.
0.1

0.000 0.01
-3.0 -2.5 -2.0 -1.5 -1.0 -0.50 0. 0.50 1.0 1.5 2.0 2.5 3.0
-4.00 -3.00 -2.00 -1.00 0.00 1.00 2.00 3.00 4.00 5.00
NSCO
NSCO

C) D)
1.20 1.20

0.80 0.80

io
V
m
ra
g io
V
m
ra
g

0.40 0.40 Azm/Plg=0/0

Azm/Plg=0/0
Model
Mean simulation Azm/Plg=90/0
Model
Mean Simul.
0.00 0.00
0. 200. 400. 600. 800. 1000. 0. 200. 400. 600. 800. 1000.
Distance Distance

FIG 6 - Simulated normal score statistics. (A) Mean of realisation histograms; (B) realisation cumulative probability curves (grey) plus mean
of the curves (red thick line); (C) variogram model (green) plus mean of realisation variograms (red dashes); (D) same as in (C) plus ten
realisation variograms along two directions (grey full and dashed lines).

TABLE 1
Simulated and interpreted fault block volume statistics. The
interpreted (model) fault block volumes have been reset to 100.

Fault Simulation volume Model Model to

block volume simulation
Min Max Average CV
% Chg
1 20 135 89 28.3% 100 -10.8%
2 57 147 108 16.8% 100 7.8%
3 33 114 86 22.2% 100 -13.9%
4 79 133 101 11.0% 100 1.4%
5 42 201 109 28.1% 100 8.6%
6 52 156 104 22.2% 100 3.6%
7 80 124 99 9.6% 100 -0.7%
≈ 100% 8 63 171 103 20.6% 100 3.2%
9 51 135 96 19.6% 100 -4.1%
≈ 95%
10 501 176 110 29.9% 100 10.1%
11 83 120 99 8.2% 100 -0.6%
≈ 67%
input data, the simulated faults are conditioned to the available
-30m -20m -10m 0m +15m +30m +45m snap points and the fault location uncertainties as assessed by the
geologist have been reproduced. The simulated fault surfaces are
visually reasonable and have more chance to be close to the
FIG 7 - Example of asymmetrical fault fluctuation distributions. interpretation than far away. On average, the interpreted fault
Two normal distributions with zero means but different standard block volumes are generally well reproduced by the simulation.
deviations are, in effect, ‘glued together’. In this example, the Where there are discrepancies, they can be explained by the
standard deviations are ten and 15 m. Simulated values outside asymmetrical input uncertainties. Although not necessary for the
the ±3 standard deviation interval are reset to three standard Resolution deposit, different aspects of the simulation could be
deviations. modified and/or improved. With the current simulation

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 179 303
G VERLY et al

Simulated fault deviations from interpretation

600 800 1000 1200 1400 1600 1800 2000 2200 2400

0 50 100 150 200 250

250 200 150 100

Deviations
50
Deviations

0 - 50 -100 -150 - 200 - 250

-250 -200 -150 -100 -50

600 800 1000 1200 1400 1600 1800 2000 2200 2400
Along strike
Interpretation (0 Deviation) Uncertainty bandwidths 20 Realisations

FIG 8 - Simulated fluctuations from one fault interpretation.

11 21 31 41

51 61 71 81

FIG 9 - Eight realisations of the fault block model at mid elevation in the model. The red lines represent the original fault interpretations.
The grey lines represent the simulated fault traces.

methodology, the total number of faults before and after
simulation is static, as per the geological interpretation. The two
semi-normal distribution rule used to rescale the fault fluctuations
could be changed to generate wider or narrower fluctuations with-
in the specified uncertainty bandwidths. The range of the Gaussian
variogram could be revisited. The simulation methodology could
be changed to generate surfaces that are more planar.
FAULTED ROCKTYPE SIMULATION
In the simulation described below, faulted rock types represent
rock types that are interpreted to have been offset by faulting.
These rock types consist of stacks of subhorizontal layers within a
background made of two rock types separated by an unconformity.
The faulted rock types are intruded by dykes and breccias that do
not appear to be offset by faulting. The general procedure for
simulating faulted rock types consists in processing one fault
block at a time. The preintrusive/prebreccia rock types are
reconstructed. The 3D interpretation of each layer is converted to a
2D interpretation of elevations and thicknesses. 2D simulations of
the elevations and thicknesses are performed, conditional to drill
hole intercepts and accounting for the uncertainties specified by
the geologist. The 2D simulation results are converted back to 3D.
The faulted blocks are stitched back together at the very end. This
simulation methodology is further described below and is
illustrated in Figure 10.
Presimulation work
The available information consists of a 3D gridded geological
model, uncertainty information and drill hole sample locations.
The uncertainty information is the same for all layers – a 95 per
cent confidence interval (‘two sigma’) of ± 30 m from the
interpreted elevation and of ± 20 per cent from the interpreted
thickness. Note the deviations for thickness are in terms of a
relative percentage of the interpreted (modelled) thickness.

304 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 FAULTED GEOLOGICAL MODEL SIMULATION OF THE RESOLUTION PORPHYRY COPPER DEPOSIT

For each fault block

Step 1 - Pre-simulation processing 3D
1a) Get the maximum simulated fault block footprint
1b) Get 3D Indicator Grid:
- Generate an indicator grid of the rock types
- Reconstruct pre-intrusive rock types
- Extend rock types from interpreted to maximum simulated fault block limit
Repeat Steps 2, 3, and 4 for each stack of layers in the fault block
Step 2 - Pre-simulation processing 2D
2a) Create 2D grids of the elevation and thickness of each layer
2b) Create 2D conditioning sample dataset from drill-hole intercepts
Step 3: Simulation 2D
3a) 2D stationary simulations of deviations, conditional to snap points.
3b) Rescale simulated deviations
3c) Incorporate rescaled deviations to interpretations.
- Adjust elevations for gaps or overlaps
Step 4: Simulation 3D
4a) Flag 3D grid from 2D simulations
Step 5: Post-simulation processing 3D
5a) Recombine the 3D grids corresponding to the stacks into one 3D grid.
Nine sets of 100 realizations (one set per fault block)
Stitch the faulted block rock types back together
For a given realization, pick the simulated rock type according to its location and
simulated fault block number at that location
100 simulated faulted rock type grids

FIG 10 - Faulted rock type simulation process.

As previously mentioned, the faulted rock types are processed
within one fault block at a time. In a first step, the
preintrusive/prebreccia faulted rock types are reconstructed
(Step 1b, Figure 10 and Figure 11). This can be done by either
using the preintrusive wireframe model, or by ‘eroding away’ the
intrusives/breccias from the grid model. The reconstructed rock
types are then extended to the maximum possible fault block
limits. Again, this could be done using the original wireframe
model or numerically approximated on the grid model. The
maximum extent of the simulated fault block is obtained by
scanning the realisations of the previously simulated fault model.
Each reconstructed and extended stack of layers is processed in
turn. The 3D interpretation of each layer is converted to a 2D
grid of elevations and thicknesses. The few drill hole intercepts –
one to five per fault block – are converted to 2D snap points.
Simulation
A series of 2D simulations are performed for each stack:
• same simulation of elevation fluctuations for all layers (will
tend to prevent crossovers), and
• one simulation of thickness fluctuations per layer.
Standard normal scores are simulated first, followed by some
rescaling to account for the uncertainty. The simulations are
conditional sequential Gaussian using a unit-sill Gaussian
variogram model with no nugget effect and a 500 m isotropic
range to ensure smooth surfaces. The checks at this stage consist
in verifying that the averages of the simulated normal scores are
close to zero. The variances of the simulated normal scores are
not expected to equal one because of the conditioning and the
relatively small extent of the fault blocks compared to the
variogram range.
Rescaling the simulated normal scores is needed to account for
the uncertainty. The assumption is that the simulated fluctuations
far from snap points are normally distributed with 95 per cent of
the values within ± the specified uncertainty. In other words, the
specified uncertainty is equivalent to two standard deviations of
the normal distribution. The uncertainty for the elevation is 30 m –
the simulated normal scores are multiplied by 15, and reset to ± 45
m if they exceed these values. The uncertainty for the thickness is
20 per cent of the interpreted thickness – the normal scores are
multiplied by ten per cent and then reset to ± 30 per cent if they
exceed these values. The final simulation step incorporates the
interpretation and simulation together. The simulated elevation at a
location is the interpreted elevation plus the rescaled simulated
deviation at that location. The simulated thickness is the
interpreted thickness plus the product of interpreted thickness and
the rescaled relative simulated deviation. Elevation corrections are
made to avoid gaps and overlaps between the simulated layers
within a same stack.
One hundred realisations of elevations and thicknesses are
available at this stage for each layer. Figure 12 shows one section
with the trace of one layer interpreted elevation and thickness
(red lines), the maximum possible deviations (blue lines) and 20
realisations of the elevation and thickness (grey lines). Three
snap points were available for the elevation but only two for the
thickness because of an incomplete hole intercept. The jagged
aspect of the thickness profiles is due to the exaggerated
resolution of the vertical grid.
Post-simulation work
One hundred realisations of one stack of layers within a fault
block are available at this stage as a series of 2D grids of layer
elevations and thicknesses. For each realisation, the layer’s top
and bottom elevations are calculated and used to flag one 3D
grid. The simulation procedure is repeated for the other stacks
within the fault block. The 3D grids are then merged together.
The procedure is repeated for the other fault blocks resulting in
11 individual simulated models. The final simulated faulted rock
type block model is obtained by merging the eleven models
together. For a given realisation and a given location, the
simulated fault block is first identified. The simulated rock type
is then picked from the rock type model that corresponds to the
fault block. Figure 13 shows eight realisations of a vertical
section through the faulted rock type model at about mid
northing in the model.
Results and discussion
Table 2 shows the statistics of the simulated faulted rock type
model. To preserve confidentiality, each of the interpreted rock

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 179 305
G VERLY et al

FIG 11 - Faulted rock types presimulation steps. (A) Original interpretation; (B) maximum simulated fault block footprint identified
on left and right; (C) preintrusive/breccia reconstruction; (D) extension to fault block maximum simulated footprint; and
(E) 2D grid of elevations and thicknesses of the mid-layer (pzls).

Mid-Layer Simulated Profiles

800 1 000 1200 1400
300
300

Elevation

Elevation
250
250
Elevation

200
200

150
75 150

800 1 000 Northing 1200 1400

75

Thickness
Thickness

Thickness
50

50
25

25

800 1000 1200 1400

Nor thing
Interpreted elevation/thickness Elevation +/- 30 m/thickness +/- 20 per cent
Snap locations (within 100 m of section) 20 Realisations

FIG 12 - Simulated elevation and thickness for one layer.

306 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 FAULTED GEOLOGICAL MODEL SIMULATION OF THE RESOLUTION PORPHYRY COPPER DEPOSIT

additional uncertainty should probably be injected in the model. A

TABLE 2 sensitivity study of the results to the range of the variogram would
Simulated and modelled faulted rock type volume statistics. be useful, as local variations could be important as the mine get
The model rock type volumes have been reset to 100. started.
Rock Simulation volume Model Model to
type
Min Max Average CV
volume simulation CONCLUSIONS
% Chg
A general method to realistically simulate the geology is
kvs 98 102 100 0.8% 100 -0.1% suggested in this paper. The method consists of first establishing a
kqs 90 113 101 4.6% 100 1.2% priority in the geological features and of simulating them one at a
pzls 89 110 99 4.7% 100 -0.7%
time. The simulation starts with the interpreted outlines or shapes
and incorporates the uncertainties as specified by the geologist.
mesc 89 114 101 4.8% 100 1.1% Two steps that could be part of such methodology were further
qzite 91 104 99 2.5% 100 -1.5% developed:
diab 97 103 100 1.4% 100 0.2% 1. simulation of a faulted block model, and
2. simulation of layered rock types within the faulted blocks.
type volumes has been reset to 100. The table shows that on The method is relatively simple because the geological features
average the interpreted (modelled) rock type volumes are are simple once they are treated one-at-a-time. For example, the
relatively well reproduced by the simulation. Differences from faults are surfaces and the faulted rock types are layers. Both
-1.5 per cent to +1.2 per cent are observed, which is considered features were simulated by simple 2D sequential Gaussian
acceptable, especially since part of these differences is related to simulations with various rescaling to account for the uncertainties
the simulated fault blocks with their asymmetric uncertainties of specified by the geologist. Last, the method is flexible in the sense
boundary positions. The table also shows the kind of differences that it can cope with complex geological models, variable
that can be observed between two realisations. For example the uncertainties and a scarcity of data. The method is made powerful
average simulated volume for pzls (Paleozoic limestone) is 99
by its incorporation of geological interpretation and uncertainty
with minimum/maximum counts of 89 and 110 respectively. The
assessments, although these can be subjective.
coefficient of variation (CV) is 4.7 per cent. Assuming a normal
distribution, there is a 95 per cent probability that the simulated
pzls volumes are within +/- 9.5 per cent of the average. These ACKNOWLEDGEMENTS
differences are significantly less than those observed for the fault
block volumes (Table 1). This is partly due to the fact that the The authors wish to thank Resolution Copper Mining, LLC and
same rock type is found in most faulted blocks. The variability of AMEC for permission to publish this paper. They also thank Harry
the rock types within any given faulted block is more significant. Parker and Geoff Ballantyne for their useful comments and
suggestions.
The approach used to simulate the faulted rock types is
considered reasonable. Indeed, the fault blocks are simulated. The
order relations between the simulated rock types are reproduced. REFERENCES
The simulated results are statistically consistent with the input Alabert, F, 1987. Stochastic imaging of spatial distributions using hard
data. The simulated rock types are conditioned to the drill hole and soft information, Master’s thesis, Department of Applied Earth
intercepts. The fluctuations in elevation and thickness reflect the Sciences, Stanford University, p 198.
uncertainty assessed by the geological staff. Still, several aspects Alfaro, M, 2008. Microergodicity and geostatistical simulation, in
of the simulation could be improved. The method assumes that Proceedings Eighth International Geostatistics Congress:
there are enough holes to get the interpretation of rock sequence Geostatistics Santiago ’08 (eds: J M Ortiz X and Emery), 1:409-417
100 per cent right at each node. This is optimistic and some (FCFM U: Chile).

11 21 31 41

51 61 71 81

C u Sh ell

FIG 13 - Simulated faulted rock type sections for eight realisations, in the northern part of the simulated area.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 179 307
G VERLY et al
Armstrong, M, Galli, A, Le Loc’h, G, Geffroy, F and Eschard, R, 2003.
Plurigaussian Simulations in Geosciences, 160 p (Springer: Berlin).
Chiles, J P, Aug, C, Guillen, A and Lees, T, 2007. Modelling the
geometry of geological units and its uncertainty in 3D from structural
data: The potential-field method, in Orebody Modelling and Strategic
Mine Planning, second edition (ed: R Dimitrakopoulos), pp 313-320
(The Australasian Institute of Mining and Metallurgy: Melbourne).
De Fouquet, C, Beucher, H, Galli, A and Ravenne, C, 1989. Conditional
simulation of random sets, application to an argilaceous sandstone
reservoir, in Proceedings Third International Geostatistics Congress:
Geostatistics Avignon ’88 (ed: M Armstrong), 2:571-530 (Kluver
Academic Publishers: Boston).
Hammitt, J and Ballantyne, G, 2007. Geologic interpretation and model
of the resolution deposit, unpublished proprietary data, Resolution
Copper Mining, LLC.
308 Spectrum Series Volume 17
Paul, A H, and Manske, S L, 2005. History of exploration at the Magma
Mine, Superior, Arizona, in Proceedings Geological Society of
Nevada Symposium 2005: Window to the World, pp 629-638 (eds:
H N Rhoden).
Srivastava, R M, 2005. Probabilistic modelling of ore lens geometry: An
alternative to deterministic wireframes, Mathematical Geology,
35(5):513-544.
Verly, G, Brisebois, K and Hart, W, 2008. Simulation of geological
uncertainty, resolution porphyry copper deposit, in Proceedings
Eighth International Geostatistics Congress: Geostatistics Santiago
’08 (eds: J M Ortiz and X Emery), 1:31-40 (FCFM U: Chile).
Advances in Orebody Modelling and Strategic Mine Planning I

Geologically Enhanced Simulation of Complex Mineral Deposits
Through High-Order Spatial Cumulants
H Mustapha1 and R Dimitrakopoulos2
ABSTRACT
Earth sciences and engineering phenomena such as geologic units, grade
content and other properties of a mineral deposit, as well as attributes of
other natural phenomena, represent complex geological systems
distributed in space. Their spatial distributions are currently predicted
from finite measurements and second-order spatial statistical models,
which are limiting, as geological systems are highly complex,
non-Gaussian and exhibit non-linear patterns of spatial connectivity.
Non-linear and non-Gaussian high-order geostatistics is a new area of
research based on higher-order spatial connectivity measures and spatial
cumulants. Key elements of a high-order spatial stochastic modelling
framework are developed herein, starting with the definitions of
high-order spatial statistics and, more specifically, the definition and
properties of spatial cumulants, and the inference and interpretation of
high-order anisotropic cumulants. Spatial cumulants are shown to capture
the directional multiple-point periodicity and spatial architecture of
geological processes. It is further shown that only a subset of all the
cumulant templates has to be computed to characterise complex spatial
patterns. The second key element of high-order geostatistics is the
simulation of complex mineral deposits using a nonparametric Legendre
series approximation with coefficient calculated in terms of spatial
cumulants. Examples show that the approach works very well.
INTRODUCTION
In earth sciences and engineering, measurements of phenomena
under study represent complex non-Gaussian systems distributed
in space. Frequently, it is required that geo-environmental
attributes are modelled and their spatial distributions predicted
from a limited set of measurements. Random field models and
stochastic data analysis, termed geostatistics, have long been
established and used as the key approach to modelling and
predicting natural phenomena in a variety of earth sciences and
engineering fields (eg David, 1977; Journel, 1989; Ripley, 1987;
Cressie, 1993; Kitanidis, 1997; Hohn, 1999; Goovaerts, 1997;
Chilès and Delfiner, 1999). Despite the considerable developments
over the past three decades, modelling approaches are based on
second-order statistics, and the spatial information these contain.
Concerns articulated during the last decade suggest that current
modelling frameworks are limited in their ability to account for
the spatial complexity of the natural phenomena being modelled,
that are critical to modelling and predicting spatially-distributed,
location-dependent data (eg Guardiano and Srivastava, 1993).
Several attempts to develop new techniques dealing with spatial
complexity include the multiple-point approach (Strebelle, 2002;
Zhang et al, 2006), new Markov random field based approaches
(Daly, 2005; Tjelmeland and Eidsvik, 2005), and others. These
developments replace the two-point covariance with a training
image so as to account for high-order dependencies. Although
these are novel approaches, there is a need for a well-defined
spatial stochastic modelling framework capable of dealing with
the complexity of geo-environmental phenomena. The approach
advocated herein is based on cumulants, which are combinations
of moment statistical parameters allowing the complete
1. COSMO – Stochastic Mine Planning Laboratory, Department of
Mining and Materials Engineering, McGill University, Montreal QC
H3A 2A7, Canada. Email: hussein.mustapha@mcgill.ca
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montreal QC H3A 2A7, Canada.
Email: Roussos.dimitrakopoulos@mcgill.ca
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
characterisation of non-Gaussian random variables. In multiple
point statistics, training images are used as a model for high order
joint distributions. However this model does not necessarily
represent the true joint distribution of the random field under
consideration. The multiple-point is a particular case of the high-
order moment and does not infer from a concrete statistical theory.
While cumulants that have statistical advanced properties
compared to the moments are explored in this paper and will be
essential in a future work. Spatial cumulants are a new concept
and it is introduced here because cumulants completely
characterise non-Gaussian stationary and ergodic spatial random
fields, thus can provide a new consistent framework in addressing
issues mentioned above. Related works on cumulants of one-
dimensional random function models have been developed to deal
with the identification, analysis and testing of non-linear signals
(eg Mendel, 1991; Nikias and Petropulu, 1993).
In the following sections, this paper first outlines basic
definitions, summarises approaches to calculating anisotropic
spatial cumulants, shows examples, and comments on relations
between cumulants and geological process. Then, the simulation
of mineral deposits with complex non-Gaussian and non-linear
geological patterns is outlined based on the use of spatial
cumulants in the high-dimensional space of Legendre polynomials
(Lebedev, 1965). Examples from simulating a complex channel
system illustrate the practical aspects of high-order simulations
based on spatial cumulants. Conclusions and comments follow.
HIGH-ORDER STATISTICS OF NON-GAUSSIAN
SPATIAL RANDOM FUNCTIONS
Let (Ω, ℑ, P) be a probability real space. A spatial real random
field Z(x) is a family of random variables {Z( x 1 ), Z( x 2 ),...} at
locations, x1,x2,..., where each random variable is defined on (Ω,
ℑ, P) and takes real values.
Assuming Z(x) is a zero-mean ergodic stationary random field,
then the cumulants of the random field Z(x) are defined from the
MacLaurin expansion of the cumulant generating function:
∼ = 1n (E{exp (jvTZ)})
ψ(v)
If its moments up to order r exist, then:
Mom[Z(x)Z(x + h1),...,Z(x + hr -1)] = E{Z(x)Z(x + h1)
... Z(x + hr -1)}
The moments depend only on the vectors h1, h2,...,hr -1.
Similarly, the rth-order cumulants of Z(x) can be denoted as:
crz ( h1 , h2 ,..., hr − 1 ) = Cum[ Z( x ), Z( x + h1 ),..., Z( x + hr − 1 )]
For example, the second-order cumulant of a non-centred
random function Z(x) known as the covariance is given by:
c2z ( h) = E {Z( x )Z( x + h)} − E {Z( x )}2
Its third-order cumulant is given by:
c3z ( h1 , h2 ) = E {Z( x )Z( x + h1 )Z( x + h2 )}
− E {Z( x )}E {Z( x + h1 )Z( x + h2 )}
− E {Z( x )}E {Z( x + h1 )Z( x + h3 )}
− E {Z( x )}E {Z( x + h2 )Z( x + h3 )} + 2 E {Z( x )}3
Spectrum Series Volume 17 309
H MUSTAPHA and R DIMITRAKOPOULOS
where:
h3 is the difference between vectors h1 and h2
It may be computationally convenient to consider zero-mean
random functions as some of the terms vanish. In addition, note
that the cumulants of orders higher than three of a zero mean
random function are related to their moments of lower orders and
a combination of their moments of order two.
Calculating experimental anisotropic spatial
cumulants
In this section, the definitions and implementation details of the
calculations of experimental cumulants from exhaustive and
sparse data sets are described.
Definitions
Spatial cumulants are defined in terms of distances in space.
Existing cumulant calculations assume regularly sampled data
and/or a regularly sampled training data set. In general, however,
geological data are available only on irregularly spaced borehole
locations. Similarly to anisotropic experimental variograms, it is
possible to restrict the calculation of cumulants to a given
direction. For this purpose, the concept of spatial template for
calculating cumulants is introduced here. A spatial template T is
defined as a particular geometry of points r r in space;
formally, given a set of directional vectors h1 , h2 ,..., hr , supported
by the direction angles {θ1,θ2,...,θr} the associated spatial
template of order (n + 1) is defined (considering a spatial
location x as a reference) as:
r r r r
Tr + 1 ( x , h1 , h2 ,..., hr , θ 1 , θ 2 ,..., θ r ) = {x , x + h1 , x + h2 ,..., x + hr}
For example, the r=3 order cumulant with the given template
T3 is computed from:
1 N N r r
C hT13 , h 2 = 2 ∑ ∑ Z( x )Z( x + hi )Z( x + hj ),
N i=1 j=1
r r
{x ; x + hi ; x + hj} ∈ T3
Figure 1 shows an example of a template associated to a third-
order experimental cumulant.
FIG 1 - L-shape template for cumulant calculation on regular
grid data.
Implementation on regular grids
In the case of regular grids, the calculations are straightforward
for particular templates that are in the same Cartesian system as
the grid. In this case, the distances and angles are defined in
terms of grid increments. For example, the template shown in
310 Spectrum Series Volume 17
Figure 1 can be easily implemented in a finite difference way. If
r
the orthonormal r of the grid is defined with the supporting
basis
unit vector i and j. For example, the experimental third-order
cumulant associated with the template T shown in Figure 1 is
given by:
1 N N r r
C hT13, h = 2 ∑ ∑ Z( x )Z( x + hi )Z( x + hj ),
2 N i=1 j=1
r r
{x ; x + hi ; x + hj} ∈ T3
Then, the algorithm is summarised as follows:
r r
1. Choose a template T by defining hi and hj . The angles are
defined by the grid: 0.45 or 90°. Start with closest grid node
(i+1,j), (i,j+1).
2. Loop until all possible triplets are multiplied.
3. Average the products.
4. Increase grid node separation and go back to step 2.
Implementation on irregular grids
For irregular grids, the situation is computationally more complex.
Computations involve angles and distances calculations, that is,
the grouping of points in terms of angles and distances. In
r more practice, the directional neighbourhood search is equivalent as
computing several directional variograms, with given angles and
ranges as shown in Figure 2.
r r
FIG 2 - Irregular template for third order cumulant calculation. h2
and h1 are the distances with Th1 and Th2, tolerance distances
respectively. a1 and a2 are the angles with tolerances Ta1 and Ta2,
respectively. (i,j) is the basis of the cartesian coordinate system.
The algorithm for third-order irregular experimental cumulant
(Figure 3) calculations (see Figure 2 for complementary
information) can be summarised as:
r r
• Choose a template by defining hi and hj . The angles a1 and a2
can be anything between 0 and 360°. Choose tolerance
angles Ta1 and Ta2 for each of the vectors. If angle equal to
90°, omnidirectional cumulants are calculated.
• Choose a tolerance distance Th1 and Th2 for each direction.
• Find the points inside the cone angle for each vector
(Figure 2).
• Calculate all the possible products between points inside
dashed boxes and supporting point at the origin {Z(x), Z(x1),
Z(x2)}.
r r
• Classify the products in terms of both distances h1 and h2
and average in every lag.
Advances in Orebody Modelling and Strategic Mine Planning I
 GEOLOGICALLY ENHANCED SIMULATION OF COMPLEX MINERAL DEPOSITS

SEQUENTIAL SIMULATION WITH HIGH-ORDER The discrete Legendre polynomials also satisfy:
SPATIAL CUMULANTS k
2
∑ P ( z )P ( z )Δz = 2 m + 1 δ
m i n i mn , ∀m, n ≥ 0 (5)
Approximation of a joint probability density using i=1

Legendre series
where:
This section discusses the approximation of continuous densities
using Legendre series. A squared integrable and real piecewise z = zi - zi-1 = 2/k is a space step
smooth function f defined on D = [-1,1] can be formally written k is the number of steps
in a series of Legendre polynomials:
{zi} is a uniform discretisation of [-1,1]
∞
P ( z) δmn is the delta Dirac function
f ( z) = ∑ Lm m
Pm
(1)
m=0 To avoid numerical instability in polynomial computation, we
normalised the Legendre polynomials by utilising the square
where Pm(z) is the mth -order Legendre polynomials, with norm norm. The set of normalised Legendre polynomials is defined as:
Pm , defined as (Lebedev, 1965): 2m + 1
Pm ( z) = Pm ( z)
m m 2
1 ⎛ d⎞
Pm ( z) = m⎜ ⎟
2 m! ⎝ dz⎠
[( z 2
]
− 1)m = ∑ ai , m zi , and − 1 ≤ z ≤ 1
i=0
(2) In this case, the orthogonality condition given in Equation 4
becomes:
The Legendre polynomials Pm(z) obey the following recursive k
relation: ∑ P ( z )P ( z )Δz = δ
m i n i mn , ∀m, n ≥ 0 (6)
i=1
2m + 1 m
Pm + 1( z) = xPm ( z) − Pm ( z) (3)
m +1 m + 1 −1 The coefficients Lm in Equation 1 of the Legendre series, the
so-called Legendre cumulants, can be determined using the
where Po(z) = 1, P1(z) = z, and m ≥1. The set of Legendre orthogonality property in Equation 4 as:
polynomials {Pm(z)}m forms a complete orthogonal basis set on
the interval [-1,1]. The orthogonality property is defined as: Lm = ∫ Pm ( z) f ( z)dz = gm ( ci ),i = 0,..., m and m = 0,1, 2,.. (7)
D

m≠n
⎧⎪ 0 , 2 where ci is the ith -order cumulant of f. Theoretically, the series 1,
∫DPm( z )Pn ( z )dx = ⎨ ,
⎪⎩ 2 m + 1 m = n
(4) with coefficients Lm calculated from Equation 7, converges to
f(z) at every continuity point of f(z) as demonstrated by Lebedev

FIG 3 - The four experimental third-order cumulant templates used in the following examples. 1- L-shape, 2 - 45°, 3-XX and 4-YY. h1 is
the distance between x and x1 and h2 is the distance between x and x2.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 311
H MUSTAPHA and R DIMITRAKOPOULOS
(1965). Finally, if only cumulants of order smaller than or equal
to ω are given, then the function f(z) in Equation 1 can be
approximated as follows:
ω
~
f ( z) ≈ fω ( z) = ∑ Lm Pm ( z) (8)
m=
The above is detailed for three-dimensional spaces in
Mustapha and Dimitrakopoulos (2010a).
` CALCULATION AND GEOLOGICAL
FIG 4 - (1) Training image; hard data locations in (2).
A high-order simulation method
This section describes the high-order conditional simulation
method (hosim) based on spatial cumulants. A sequential
procedure simulating values at unsampled locations that are
randomly visited is used here. The Legendre series
approximation is used to estimate the cpdfs (Mustapha and
Dimitrakopoulos, 2010a). This expression uses Legendre
polynomials which are orthogonal on the finite interval [-1,1].
Then, the training images (Figure 4 (1)) and the data values
(Figure 4 (2)) are first scaled to [-1,1]d, where d is the dimension
of the problem (ie, d = 1,2 or 3).
The hosim method first combines the training image (TI) used
and the samples (Figure 5) to infer the high-order spatial
cumulants. A global calculation procedure is performed based on
a given maximal template size (TEMP) (Figure 5 (2)). This step
consists of calculating all the spatial cumulants needed by the
Legendre series approximation.
The main steps of hosim method are as follows:
1. Scan the training image and the sample data (Figure 5 (1))
and store the spatial cumulants in a global tree.
2. Define a random path visiting once all unsampled nodes.
(1)- Training image with hard data (2)- Template TEMP for
global calculation
FIG 5 - Training image in (1). The template in (2) is used for a
global calculation of the spatial cumulants.
312 Spectrum Series Volume 17
3. Define the template shape T for each unsampled location x0
using its neighbours. The conditioning data available within
TEMP are then searched (Figure 5). The high-order spatial
cumulants are read from the global tree in step 1,and are
used to calculate the coefficients of the Legendre series.
These coefficient are used to build the cpdf of Z0.
4. Draw a uniform random value in [0,1] to read from the
conditional distribution a simulated value, Z(x0), at x0.
5. Add x0 to the set of sample hard data and the previously
simulated values.
6. Repeat steps 4 and 6 for the next points in the random path
defined in step 3.
7. Repeat steps 3 to 7 to generate different realisations using
different random paths.
The random path defined in step 5 concerns only the
unsampled locations. Thus, the final realisation obtained in after
step 8 honours the conditioning data.
SPATIAL CUMULANTS – EXAMPLES,
INTERPRETATIONS
Examples of third-order cumulants calculated on two- and
three-dimensional data sets are presented in this section. Results
are interpreted so as to understand their use as a pattern
recognition tool. The data sets utilised represent complete images
and the regular grid approach described in a previous section is
followed. It should be noted that, the covariance is a measure of
the periodicity between pairs of points separated by given
distances; similarly, the higher-order cumulant is also a measure
of periodicity but in the direction of the symmetry axis of
template used, that is, the multiple point symmetry. In the
examples that following, cumulants are computed on zero mean
data sets.
Spatial templates
In the following example, the covariance map and four
directional experimental cumulants are presented, unless
otherwise specified. Fourth- and fifth-order cumulants are,
respectively, computed with the and xy(-x)(-y) templates as
shown in Figure 6.
Training image
In Figure 7 (2) and Figure 7 (3) we consider an interpretation of
a diamond baring kimberlite pipe of the Ekati mine, NT Canada
(Nowicki et al, 2004), and its translation to a 3D binary training
image (76 055 nodes are used). The geological interpretation of
the pipe geometry is seen as two parts: one on the top and
another one on the bottom as shown in Figure 7. The part on the
bottom is close to a three-dimensional cone geometric shape.
The base of this cone or the intersection between the two
parts of the pipe is a horizontal section of size 150 m along
x and 200 m along y (Figure 7 and Figure 11). (For more
details, see Dimitrakopoulos, Mustapha and Gloaguen, 2010 and
Mustapha and Dimitrakopoulos, 2010b). Results for high-order
cumulants are presented first. Then, a discussion follows.
The third-order cumulant maps provide approximations of the
pipe shape as shown in Figure 8 (1) to Figure 8 (3). The fourth-
order xyz cumulants average the objects and translate them to the
origin.
For example, the pipe, considered as the only object in
Figure 7 (3), is translated to the origin as shown in Figure 9 (1).
From this figure, 2D cross-sections are shown in Figure 9 (2) and
Figure 9 (3). These cross-sections provide, approximately, the
Advances in Orebody Modelling and Strategic Mine Planning I
 GEOLOGICALLY ENHANCED SIMULATION OF COMPLEX MINERAL DEPOSITS

1- xy(-y) 2- xy(-x)(-y)

FIG 6 - Examples of fourth-order cumulant templates (1) and fifth-order cumulant templates (2).

(2) Geological (2) Translation of the pipe

(1) Drill holes
interpretation of (1) (2) to a 3D training image

FIG 7 - A geological interpretation of a kimberlite pipe, (1) drill holes; (2) a geological interpretation of (2); (3) regular block approximation of
the pipe surrounding rock.

results obtained by the third-order cumulant maps in Figure 8.
This conclusion is justified by the fact that the pipe shape is
approximately reflected, in a specific plane, by using orthogonal
shapes (xy or L-shape, xz or yz) cumulants, while it is fully
characterised using xyz cumulants. These observations show, not
surprisingly, the ability of the higher order cumulants to include
key relations seen in lower orders. More generally, relations
between cumulants can be extended for an order higher than four
and, in particularly, order five. The fifth-order cumulant maps are
based on four directions, and placed in four-dimensional space.
Then, cross-sections are used as detailed before. For example,
Figure 10 shows a 3D cross-section of the xyz(-z) fifth-order
cumulant map. This figure translates the pipe to the origin and
reflects the results of the fourth-order and, consequently, the
third-order results are reflected too. The pipe shape varies
strongly between the bottom and the top along the vertical axis
(z) while the variations are less along the horizontal axis (x and y)
as shown in Figure 11. Figure 12 shows results of the third-,
fourth- and fifth-order with a different way. Figure 12-(1.a) to
Figure 12-(3.a) present third-order cumulant maps using lines
contours from Figure 8.
Several 2D cross-sections, along x, y and z directions, are
selected from the fourth-order cumulant maps in Figure 9. The
fronts from 2D cross-sections of Figure 10 are selected as shown
in Figure 12 (1.c) to Figure 12 (3.c). The third-order cumulant
map in Figure 12 (1.a) shows a regular shape of the horizontal
sections of the pipe while the fourth- and fifth-order cumulant
maps, in Figure 12 (1.b) to Figure 12 (1.c), reflect some
horizontal irregularity of the pipe. The main raison comes from
the variation of the pipe size along the vertical axis.
This variation is more described by the fourth- and the fifth-
order cumulants because they manipulate points in more than two
directions. For example, four points are considered for the fourth
order-cumulant and one of the points varies along the z direction.
The variations along x and y axis are less than the variation along
z axis. The third-order cumulant maps in Figure 12 (2.a) to
Figure 12 (3.a) show, approximately, results close to those
obtained from fourth- and fifth-order maps as shown in Figure 12
(2) and Figure 12 (3). The fourth- and fifth-order maps detect
with more precision the size along x (150 m) and y (200 m) of
the intersection between the two parts of the pipe.
Drill hole data
The cumulant maps calculated on the 3D training image provide
good interpretation for the pipe’s shape. In the following,

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 313
H MUSTAPHA and R DIMITRAKOPOULOS

(1) Third-order: xy- (3) Third-order: yz-

(2) Third-order: xz-templates
templates templates

FIG 8 - Third-order cumulants. (1) to (3) are third-order cumulants for the Fox kimberlite pipe, NT in Figure 7 (3), using xy, xz and yz
templates.

(2) Fourth-order : Three 2D cross-sections

(1) Fourth-order: xyz- templates
from (1)

FIG 9 - Fourth-order cumulants. (1) is a fourth-order cumulant map for the Fox kimberlite pipe, NT in Figure 7 (3), using xyz templates; (2)
are three 2D cross-sections from (1).

cumulants are calculated on the original data obtained from the
pipe drill holes in Figure 7 (1). Figure 13 to Figure 15 show the
obtained results. In these figures, the red lines represent the
borders of the set of high values in the cumulants maps. These
borders reflect the shape of the pipe along x and y directions
which is approximately similar to the results obtained on the 3D
training image (Figure 12). The top part of the pipe is better
detected then the bottom part and that because most of the data
points are in the top part, until 300 m depth, as the pipe drill holes
show in Figure 7 (1). In Figure 13, the fourth- and fifth-order
cumulants maps calculated on data provide better description of
the horizontal sections than the third-order map. They show
irregularity between 400< × <500 and along y as detected with the
fourth- and fifth-order maps from the 3D training image.

314 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 GEOLOGICALLY ENHANCED SIMULATION OF COMPLEX MINERAL DEPOSITS

(1) Fifth-order: A 3D cross-section from the (2) Fifth-order: Three 2D cross-sections from
xyz(-z) templates cumulant map at ‘-z=0’ (1)

FIG 10 - Fifth-order cumulants. (1) is a 3D cross-section at –z=0 of the five-dimensional xyz(-z) template cumulant map; (2) are three 2D
cross-sections from (1).

(1.a) xy cross-section at z=350 (1.b) xy cross-section at z=250

(2.a) xz cross-section at y=300 (2.b) xz cross-section at y=200

(3.a) yz cross-section at x=300 (3.b) yz cross-section at z=200

FIG 11 - Different horizontal and vertical 2D cross-sections of the pipe in Figure 7.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 315
H MUSTAPHA and R DIMITRAKOPOULOS

THIRD-ORDER FOURTH-ORDER FIFTH-ORDER

(1.b) 2D cross-sections (1.c) 2D cross-sections along z
r r
(1.a) {x , y} cumulant map along z from the map in from the map in Figure 10 -(1)
Figure 9 -(1)

(2.b) 2D cross-sections (2.c) 2D cross-sections along y

r r
(2.a) {x, z} cumulant map along y from the map in from the map in Figure 10 -(1)
Figure 9 -(1)

(3.b) 2D cross-sections 2D cross-sections along x from

r r
(3.a) { y, z} cumulant map along x from the map in the map in Figure 10 -(1)
Figure 9-(1)

FIG 12 - Third-, fourth- and fifth-order cumulant maps for the pipe (3D training image) in Figure 7.

316 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 GEOLOGICALLY ENHANCED SIMULATION OF COMPLEX MINERAL DEPOSITS

xy-view
(1) Third-order (2) Fourth-order (3) Fifth-order

FIG 13 - Xy views of the third-, fourth- and fifth-order cumulant maps calculated using the data from drill holes in Figure 7 (1).

xz-view
(1) Third-order (2) Fourth-order (3) Fifth-order

FIG 14 - Xz views of the third-, fourth- and fifth-order cumulant maps calculated using the data from drill holes in Figure 7 (1).

CONDITIONAL SIMULATION USING CUMULANTS
AND TRAINING IMAGES
In this section, the simulation of a two-dimensional exhaustive
image (Figure 16 (1)) is shown so as to illustrate the high-order
conditional simulation using spatial cumulants. The example
presented herein uses 25 sample data (Figure 16 (2)), the training
image in Figure 16 (3) and the conditional simulation algorithm
discussed above.
Different realisations are presented as shown in Figure 17.
This figure shows that the main characteristics of the exhaustive
image are reproduced using a sparse data set (about 0.85 per cent
of the total number of points). The 2D sections presented here
have particular and complex distributions as shown by the
bimodal histogram in Figure 18 (1). This figure shows the
comparison between the generated realisations histograms and
the data set histogram. In addition, the realisations reproduced
the variograms along the EW and NS directions of the data set as
shown in Figures 18 (2) and (3). The developed method is also
validated by comparing the high-order statistics of the data set,
exhaustive image and the different realisations obtained. For
example, the third-order spatial cumulant maps of the exhaustive
image, data set, realisations 1 and 2 are very close as shown in
Figure 19. This last result is obtained because the new
conditional simulation algorithm uses different cumulants orders
in the Legendre series and this will guaranty the reproduction of
not only the histogram and variograms of the sample data, but
also their high-order statistics.
CONCLUSIONS
This paper presented developments towards a new alternative
approach to modelling complex, non-linear, non-Gaussian earth
sciences and engineering data, as required in most applications.
The new alternative framework is founded upon concepts from
high-order statistics that are introduced herein in a spatial context.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 317
H MUSTAPHA and R DIMITRAKOPOULOS

yz-view
(1) Third-order (2) Fourth-order (3) Fifth-order

FIG 15 - Yz views of the third-, fourth- and fifth-order cumulant maps calculated using the data from drill holes in Figure 7 (1).

(1)- True image (2)- 25 Sample data (3)- Training image

FIG 16 - Simulation of a horizontal 2D section of a fluvial reservoir. (1) Exhaustive image: true image, (2) 25 sample data, and (3) a
training image.

(1)- Realisation 1 (2)- Realisation 2

FIG 17 - Realisations (1) and (2) obtained by the hosim.

318 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 GEOLOGICALLY ENHANCED SIMULATION OF COMPLEX MINERAL DEPOSITS

(1)- Histograms (2)- Variograms NS (3)- Variograms EW

γ γ

FIG 18 - Histograms (1), NS (2) and EW (3) variograms of 10 hosim realisations. The cercles refer to the data set 1 and the solid lines refer
to the realisations.

(1)- realisation 1 (2)- realisation 2

0.01 0.01
0.015 0.015

0.02 0.02

FIG 19 - Third-order spatial cumulant maps of the realisations 1 and 2, respectively.

Mathematical definitions of non-Gaussian spatial random REFERENCES

functions and their high-order spatial statistics are described in
detail, stressing the notion of spatial cumulants. The calculation of
spatial cumulants, including anisotropic experimental cumulant
calculations using spatial templates are introduced, and examples
of three-dimensional images presented and their characteristics
analysed to assess the relations between cumulants and geological
patterns. The simulation of a complex image of channels is
presented using a new high-order sequential simulation method
which is based on the concept of high-order spatial. The results
showed a good reproduction of the main features of the exhaustive
image using a small data set. The realisations generated
reproduced the histogram, variogram and high-order statistics of
the data set. A key aspect of the simulation method based on
spatial cumulants is the compliance of the simulated realisations
with all statistics (any order) of the available data and avoid
possible conflicts between training images and dense data sets
commonly available in mining studies.
ACKNOWLEDGEMENTS
The work in this paper was funded from NSERC CDR Grant
335696 and BHP Billiton, as well NSERC Discovery Grant
239019. Thanks are in order to Brian Baird, Peter Stone, and
Gavin Yates, as well as BHP Billiton Diamonds and, in
particular, Darren Dyck, for their support, collaboration, data
from Ekati Mine, and technical comments.
Billinger, D R and Rosenblatt, M, 1966. Asymptotic theory of kth-order
spectra, Spectral Analysis of Time Series (ed: B Harris), pp 189-232
(John Wiley: New York).
Bilodeau, M and Brenner, D, 1999. Theory of multivariate statistics
(Springer: Berlin).
Chilès, J P and Delfiner, P, 1999. Geostatistics: Modeling Spatial
Uncertainty (John Wiley & Sons, Inc: New York).
Cox, D R, 1972. Regression models and life-tables (with discussion),
Journal of the Royal Statistical Society, Series B (Methodological),
34(2)187-220.
Cressie, N A, 1993. Statistics for Spatial Data (John Wiley: New York).
Daly, C, 2004. Higher order models using entropy, Markov random fields
and sequential simulation, in Geostatistics Banff, pp 215-225
(Springer).
David, M, 1977. Geostatistical Ore Reserve Estimation (Elesevier:
Amsterdam).
David, M, 1988. Handbook of Applied Advanced Geostatistical Ore
Reserve Estimation (Elsevier: Amsterdam).
Dimitrakopoulos, R, Mustapha, H and Gloaguen, E, 2010. High-order
statistics of spatial random fields: Exploring spatial cumulants for
modelling complex, non-Gaussian and non-linear phenomena,
Mathematical Geosciences, 42(1):65-97.
Goovaerts, P, 1998. Geostatistics for Natural Resources Evaluation
(Oxford: New York).
Guardiano, J and Srivastava, R M, 1993. Multivariate geostatistics:
Beyond bivariate moments, in Geosatistics Tróia ’92 (ed: A Soares)
volume 1, pp 133-144 (Kluver: Dordrecht).

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 319
H MUSTAPHA and R DIMITRAKOPOULOS
Hohn, M E, 1999. Geostatistics and Petroleum Geology (Kluwer: New
York).
Journel, A G, 1989. Five lessons in geostatistics, AGU, San Francisco.
Journel, A G, 1994. Modeling uncertainty: Some conceptual thoughts,
Geostatistics for the Next Century (ed: R Dimitrakopoulos) (Kluwer:
Dordrecht).
Journel, A G, 1997. Deterministic geostatistics: A new visit, in
Geostatistics Woolongong ‘96 (eds: E Baafy and N Shofield)
pp 213-224 (Kluwer: Dordrecht).
Journel, A G, 2000. Stochastic imaging of channels from seismic data
using mp statistics, SCRF, Report 13, Stanford University.
Journel, A G and Huijbregts, C J, 1978. Mining Geostatistics, 600 p
(Academic Press Inc: New York).
Kitanidis, P K, 1997. Introduction to Geostatistics – Applications in
Hydrogeology, (Cambridge University Press: New York).
Lebedev, N N, 1965. Special Functions and Their Applications, 308 p
(Prentice-Hall Inc: New York).
Matérn, B, 1960. Spatial variation – Stochastic models and their
application to some problems in forest surveys and other sampling
investigations, Meddelanden fran Statens Skogsforskningsinstitud,
49(5), Almaenna, Stockolm.
Matheron, G, 1971. The theory of regionalized variables and its
applications, Cahier du Centre de Morphologie Mathematique,
number 5, Fointanebleau.
Mendel, J M, 1991. Use of high-order statistics (spectra) in signal
processing and systems theory: Theoretical results and some
applications, IEEE Proceedings, 79:279-305.
Mustapha, H and Dimitrakopoulos, R, 2010a. High-order stochastic
simulations for complex non-Gaussian and non-linear geological
patterns, Mathematical Geosciences, 42(5)457-485.
320 Spectrum Series Volume 17
Mustapha, H and Dimitrakopoulos, R, 2010b. A new approach for
geological pattern recognition using high-order spatial cumulants,
Computers and Geosciences, 36(3), pp 313-334.
Nikias, C L and Petropulu, A P, 1993. Higher–Order Spectra Analysis: A
Nonlinear Signal Processing Framework, 538 p (PTR Prentice Hall:
Upper Saddle River).
Nowicki, T, Crawford, B, Dyck, D, Carlson, J, McElroy, R, Oshust, P and
Helmstaedt, H, 2004. The geology of the kimberlite pipes of the
Ekati property, Northwest Territories, Canada, Lithos. 76:1-27.
Ripley, B D, 1987. Stochastic Simulation (John Wiley & Sons Inc: New
York).
Rosenblatt, M, 1985. Stationary Sequences and Random Fields
(Birkhaüser: Boston), 258 p.
Sesay, S A and Rao, T, 1991. Difference equations for higher-order
moments and cumulants for bilinear time series models, Jour Time
Ser Anals, 12:159-177.
Strebelle, S, 2002. Conditional simulation of complex geological
structures using multiple point statistics, Mathematical Geology,
34:1-22.
Swami, A, Giannakis, G B and Mendel, J M, 1990. Linear modeling of
multidimensional non-Gaussian processes using cumulants,
Multidimensional Systems and Signal Processing, 1:11-37.
Tjelmeland, H, 1998. Markov random fields with higher order
interactions, Sacndinavian Journal of Statistics, 25:415-433.
Tjelmeland, H and Eidsvik, J, 2004. Directional Metropolis: Hastings
updates for posteriors with nonlinear likelihoods, in Geostatistics
Banff, pp 95-104 (Springer).
Zhang, T, Switzer, P and Journel, A G, 2006. Filter-based classification
of training image patterns for spatial simulation, Mathematical
Geosciences, 38:63-80
Advances in Orebody Modelling and Strategic Mine Planning I

Stochastic Mine Planning – Example and Value from Integrating
Long- and Short-Term Mine Planning Through Simulated Grade
Control, Sunrise Dam, Western Australia
A Jewbali1 and R Dimitrakopoulos2
ABSTRACT
A new multistage stochastic mine production scheduling approach is
developed and tested in a large operating gold mine. The proposed
approach takes short-scale orebody information in the form of grade
control data into account. As simulated orebodies used in stochastic
long-term mine planning are based on sparse exploration data and while
grade control data are unavailable at the time of production scheduling,
the short-scale information is first simulated stochastically and then
serves as input to the optimisation process. Stage 1 of the approach
generates high density future grade control data for incorporation into the
production scheduling process based on sequential co-simulation and
pseudo cross-variograms between exploration data and grade control in
previously mined out parts of a deposit. In Stage 2, the technique of
conditional simulation by successive residuals enables pre-existing
simulated orebody models to be updated using the simulated future grade
control information. Stage is based on a stochastic programming mine
scheduling formulation that handles multiple simulated orebody models
from Stage 2 and accommodates both maximising Net Present Value
(NPV) and minimising deviations from production targets. Stage 4
includes quantification of risk in the produced schedules generated,
comparison of schedules and reporting.
The application at a large operating gold mine demonstrates that the
proposed approach is practical and adds value to the operation. The
approach is shown to deliver additional ore (3.6 Mt more) and metal
(2.6 million grams) which matches the mined reconciliations and results
in a cumulative NPV which is on average A$7.7 M higher than that of a
stochastic schedule without the simulated grade control data and
substantially higher (about 30 per cent) compared to the NPV from the
actual schedule of the mine.
INTRODUCTION
Stochastic mine planning is a relatively recent development aimed
at addressing uncertainty in ore supply from an orebody,
commodity prices and metal demand, as well as expanding to
other issues of uncertainty in mine planning. The existing work
focuses largely on open pit mine design and production scheduling
with uncertain grades/metal content and geological conditions. It
includes life-of-mine (LOM) production scheduling based on
optimal mining rates and simulated annealing (Godoy and
Dimitrakopoulos, 2004), simulated annealing with constant
mining rates (Leite and Dimitrakopoulos, 2007), stochastic integer
programming formulations including a stochastic stock pile
(Ramazan and Dimitrakopoulos, 2009, 2007), further tests of this
last approach (Leite and Dimitrakopoulos, 2010, in this volume),
expansion for multiple elements (Benndorf and Dimitrakopoulos,
2010, in this volume) and others. The two key aspects of all the
above mentioned approaches and related example case studies are
that the derived long-term production schedules:
1. have substantially higher present value than the tradition-
ally used approaches with differences between 20 per cent
and 30 per cent, and
1. GAusIMM, Resource Estimation Geologist, Rio Tinto, 152-158
St Georges Terrace, Perth WA 6000. Email: Arja.Jewbali@riotinto.com
2. FAusIMM, Professor and Director, COSMO – Stochastic Mine
Planning Laboratory, Department of Mining and Materials
Engineering, McGill University, Montreal QC H3A 2A7, Canada.
Email: roussos.dimitrakopoulos@mcgill.ca
Advances in Orebody Modelling and Strategic Mine Planning I
HOME
2. minimise the potential deviations from production targets
set given the available drilling and orebody models.
Recent work further considers integrating market uncertainty
in choosing from different mine designs which also consider
geologic uncertainty and show further monetary benefits of
quantifying and integrating uncertainty (Dimitrakopoulos and
Abdel Sabour, 2007). Meagher, Abdel Sabour and Dimitra-
kopoulos (2010, in this volume) explore the effects of optimising
pits in a new approach that assesses the combined effect of metal
and price uncertainties using a network flow approach; their study
shows once again the value of stochastic approaches and the
integration of key uncertainties to mine planning.
The work above is all based on two elements, in addition to new
stochastic mine design and scheduling methods. Firstly, the ability
to simulate sets of equally possible representations of the orebody
being assessed. This set of orebodies is the input to the above
mentioned stochastic optimisers and represents the geological
uncertainty and local grade variability of the orebody, as
understood from the available drill hole data – this drastically
differs from the conventional single estimated (smoothed out or
average type) input to conventional mine design and optimisation
methods and is the major reason for improved results from the
stochastic scheduling methods discussed. Secondly, the ability to
assess cash flows using evaluation methods based on multiple
simulated realisations of the related commodity price, rather than
single average and typically constant, over the LOM and metal
price forecasts. Both this second element and the use of
uncertainty in metal content leads to improvements in the last two
studies above.
A limit of both stochastic and conventional long-term planning
and optimisation of production planning is that short-term
schedules may frequently deviate from the expectations of
long-term ones. While long-term production scheduling is used to
maximise the Net Present Value (NPV) of a project, short-term
production scheduling focuses on meeting short-term production
demand within the long-term plan, given the processing capacity
of the mill and managing the quality of ore being processed. Both
short- and long-term scheduling are usually based on
exploration-scale orebody models, whereas the actual short-term
production performance is controlled by local, blasthole-scale data
used for grade control and ore/waste selection, usually not
available at the time of scheduling. As a result, short-term
production sequences deviate from the long-term plans and are
adjusted to meet mill demand or the production performance may
substantially deviate from forecasts. In short, discrepancies occur
between what is forecasted and what is actually mined.
A solution to the above is to consider production scheduling
approaches that are capable of reflecting short-scale behavior of
the orebody beforehand, that is, at the planning stage and through
grade information available before the actual grade control
drilling. High density grade control information offers tighter
controls on geology and mineralisation characteristics and predicts
qualities and quantities closer to what is actually being mined.
Although not available prior to blasthole drilling, grade control
data can be simulated in different ways. Simple random sampling
errors added to simulated realisations from exploration data, have
been used to assess production schedules reported in a feasibility
study (Guardiano, Parker and Isaaks, 1995) and a conceptually
Spectrum Series Volume 17 321
A JEWBALI and R DIMITRAKOPOULOS

similar more sophisticated approach is considered for assessing As a result of the ability to link information at different scales
mineral resources by Journel and Kyriakidis (2004). In operating in space and time, short scale variability, short- and long-term
mines, errors from field duplicates or from nearest paired grade schedules are integrated. The proposed approach or process has
control holes and drill holes have been used to simulate four stages shown in Figure 1, which also shows how this
realisations of future grade control data and assess short-term compares to the four steps followed in the existing stochastic
production schedules at Escondida copper mine, Chile LOM planning:
(Khosrowshahi, Shaw and Yeats, 2007). Multiple simulations of
correlated sampling errors using data from previously mined parts 1. Stage 1: Generate high density future grade control data for
of the deposit have been used at Morila gold deposit, Mali (Peattie incorporation into the production scheduling process. This
and Dimitrakopoulos, 2009) with excellent results in data is not yet available at the time of planning and is
reconciliations. With the exception of the last one, past approaches simulated.
are relatively simplistic. All existing work in this area assesses 2. Stage 2: Update pre-existing simulated orebody models
potential risks in production schedules or potential reserves – using the simulated future grade control information and a
however, it does not address the major question: How to improve suitable stochastic simulation technique.
or reduce deviations of forecasts from actual production at the
time of planning while maximising the present value of an asset. 3. Stage 3: Generate production schedules using a stochastic
This can be effectively addressed only by technically integrating optimisation method handling multiple simulated orebody
short-scale orebody variability and grade control with, short- and models from Stage 2, which accommodates both maximising
long- term LOM production planning into one approach. net present value and minimising deviations from production
Short-scale information is the core element in the targets.
above-mentioned integration. As this information is unavailable 4. Stage 4: Quantify grade risk in the produced schedules,
and needs to be stochastically simulated at the time of LOM compare schedules with and without simulated short scale
planning, the way to utilise a range of possible scenarios is orebody information, and report mine schedules as needed.
through the stochastic optimisation formulations used in
These four stages are discussed next in greater detail.
stochastic mine planning as reviewed above. This paper
contributes a new multistage approach to production scheduling
that incorporates, short-scale deposit information and related Existing Stochastic LOM Proposed Multistage Approach with
grade uncertainty into the scheduling process. The approach Scheduling Process Short-scale Information
allows for the realistic integration of short- and long-term mine
production schedules, as well as the generation of more reliable
mine production forecasts. A case study at an operating gold Stage 1
deposit demonstrates the approach, provides comparisons Simulation of orebody models from Simulation of high density future
exploration data grade control data
between the traditional schedules of the mine, accounts for
stochastic schedules before and after simulated dense grade
control information and quantifies the expected monetary value Stage 2
of the method. In the next sections, the proposed approach is first Updating of the existing orebody
models with the future data
outlined, the case study presented, results are analysed and
conclusions follow.
Stage 3
A NEW MULTISTAGE APPROACH TO Stochastic optimisation and Stochastic optimisation and generation
PRODUCTION SCHEDULING generation of production schedules of production schedules

The multistage approach to production scheduling proposed here

is as sequence of steps employing separate techniques for each Stage 4
stage. The approach differs conceptually from the traditional Schedules
approaches in many aspects. A major difference is that it requires Quantification of risk and analysis in schedule
multiple, equally probable representations of the orebody, which
are generated from spatial Monte Carlo simulation methods and
at two different levels. The first level is that of exploration-type
data sets and information; the second level is that of grade
control data. This simulation-based framework assists in
quantifying and generating risk managing schedules as well as
accommodates the interaction of information at different scales,
both in space and time. Different scales in space refer to the:
FIG 1 - Schematic presentation of the proposed multistage mine
• local variability of the deposit and local classification of scheduling approach.
materials selected as ore and waste through a grade control
process, and
Stage 1 – simulation of future grade control
• physical parts of the orebody that are mined over long
periods of time as individual pit units (cut-backs) and parts of
drilling data
individual pit units that correspond to the parts of cut-backs High density future grade control data can be simulated using
mined over short periods of time from the same mining front. several approaches, such as those reviewed in the introduction. A
Different scales in time refer to the: more elaborate and effective method is outlined here and is used
in the case study that follows. The method, schematically shown
• yearly time units on which the long-term plans are reported, in Figure 2, assumes that there are mined out parts of an orebody
and with grade control information of comparable quality and
• relatively shorter time periods in short-term planning characteristics to the remainder (as yet un-mined) of the orebody
(weekly, monthly or quarterly). being studied.

322 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 STOCHASTIC MINE PLANNING – EXAMPLE AND VALUE FROM INTEGRATING LONG- AND SHORT-TERM MINE PLANNING
Bench/section of pit already mined out
Exploration data
Grade control data
Define relationship
Exploration data
Simulate grade control data
Bench/Section of pit NOT yet mined out
FIG 2 - Simulating future grade control data.
Exploration drilling and grade control data from a mined out
section or bench of a deposit is used to quantify the spatial
relationship between the two types of data. As it is not physically
possible to have exploration and grade control information
available at the same location, their spatial cross-correlation
(relationship) is quantified through the so-called pseudo
cross-variogram (Myers, 1991). Having quantified this spatial
relationship, a method such as the sequential Gaussian
co-simulation of two variables (Goovaerts, 1997) can be used to
generate the future grade control data in un-mined sections of an
open pit. The co-simulation approach ensures that the simulated
future grade control data have the same distribution as past grade
control information and the same auto- and cross-correlation
(with exploration data).
Stage 2 – updating of existing simulations with
future grade control data
The second stage of the production scheduling approach presented
herein involves the updating of pre-existing simulated orebody
models. This is because orebodies under study have frequently
already been simulated for other purposes, or because there can be
more than one scenario of possible sets of future data to consider,
and so on. Generally the ability to update pre- existing realisations
of an orebody is useful in improving the efficiency of the process.
The only known simulation technique that is capable of providing
updating capabilities to accommodate the updating of pre-existing
simulations with new data such as grade control is the technique of
conditional simulation by successive residuals or CSSR
(Vargas-Guzman and Dimitrakopoulos, 2002). The end result of
the use of CSSR is a set of simulated orebody models that are
conditioned on both the known exploration data and the simulated
future grade control data. Note that new information for updating
may also be additional data from near mine exploration drilling, or
in fill drilling.
Stage 3 – stochastic integer programming
formulation
To generate production schedules, a stochastic optimisation
formulation based on Stochastic Integer Programming (SIP) is
used, and it is a simpler version of the one developed in Ramazan
and Dimitrakopoulos (2007, 2010) as well as discussed in Leite
and Dimitrakopoulos (2010, in this volume). The objective
function of this formulation maximises the expected discounted
cash flows while minimising the cost of deviating from the set
production targets and it is:
P N S P R has produced more than two million ounces of gold. The pit is
Max ∑ ∑ E(NPV)ti Xti − ∑ ∑ ∑ ( s Cu trYu tr + s Cl trYl tr )
t = 1 i= 1 s = 1 t = 1 r= 1
Advances in Orebody Modelling and Strategic Mine Planning I
where:
P is the number of periods to schedule
N is the total number of blocks to schedule
S refers to the number of simulated orebody models
R is the number of targets
t
X i is a binary variable indicating block i mined in period
t (1 = mined/0 = not mined)
E(NPV)ti is the expected NPV to be generated if block i is
mined in period t
Yu tr is the excess amount produced compared to the target
(grade/tonnage) r in period t for orebody model s
s
Cu tr is the unit cost to penalise Yu tr in period t
YItr is the deficient amount produced in relation to target
(grade/tonnage) r in period t for orebody model s
s
CItr is the unit cost to penalise YItr in period t
The first component in Equation 1 maximises the total
discounted cash flow and the expected Net Present Value (NPV)
is determined by calculating economic values for each simulated
model, then averaging them. This component reflects the
conventional goal of optimising total discounted cash flow over
the LOM. The second component minimises the deviations
between expected productions as these are described by
simulated orebody models used as input and a set of mine
production targets and minimises the risk of not meeting targets
due to grade uncertainty. The deviations are calculated over the
set of simulated orebody models and are penalised with costs.
This second component reflects the goal of short-term
production scheduling, and its use together with the first
component provides the means to integrate long- and short- term
planning, in the context of the approach proposed herein. It
should be noted that a novelty of this formulation is that the costs
can be discounted for each subsequent period using a geological
discount rate. As a result, the cost related to not meeting mill
requirements can be set higher in the first period than in
subsequent periods, allowing the risk distribution between
production periods to be managed.
Stage 4 – quantification of uncertainty
Stage 4 aims to quantify grade risk in the produced schedules
that have been generated. Any schedule can be assessed against
possible scenarios of orebodies, be they constructed from
exploration datasets only or in combination with future grade
control data. This allows comparisons and assessments, including
an evaluation of effects from potential short-scale orebody
variability.
CASE STUDY AT A GOLD MINE
The gold deposit and mine
The gold deposit in this case study lies along a shear zone and
the general trend of the gold mineralisation is parallel to this
zone. Gold mineralisation can be both secondary and primary.
Secondary mineralisation is related to iron oxidation fronts and
water tables and to aggregation within paleochannels, while
primary mineralisation can roughly be related to high strain shear
zones by low strain stockwork-vein zones. Based on the deposit
geology, 17 domains are inside the open pit (sections of which
are shown in Figure 3). Mining is done using four excavators
which load into 25 rear dump trucks. Since opening, the mine
(1) about 2.1 km long and 1.1 km wide with a final depth of 450 m
to be reached at end of its life. The total amount of waste mined
Spectrum Series Volume 17 323
A JEWBALI and R DIMITRAKOPOULOS

Generation of future grade control data

2200 Within the mined out section of the pit, 75 789 grade control drill
North
P1 holes were available. These were separated based on the geological
domain they belonged to. In addition, 3934 exploration diamond
and reverse circulation holes were available within the same section
of the pit. These were also divided based on geological domains.
P2 For each geological domain these two sets of information were used
P3 to derive the relationship between exploration and grade control
data. This relationship can than be extended to unmined sections of
P4 the pit where only exploration data is available. For the purposes of
2000 this study, twenty future grade control data simulations were
Y=100089 generated, using the sequential cosimulation method discussed
earlier, on 5 m × 7 m × 2.5 m spacing in the unmined sections of
the pit (the volume displayed in Figure 3).

2200 P1
Updating existing simulations with future grade
control data
Stage 2 of the production scheduling approach involves using
P2 conditional simulation by successive residuals to update pre-
P3 existing orebody models. The 20 available simulations (on 2 m ×
2 m × 2.5 m spacing) conditioned on exploration data were
updated with the 20 simulated future grade control data generating
P4 a set of 400 simulations that are conditioned on both the known
Y=99824 exploration data and the simulated future grade control data
2000 scenarios. Twenty updated simulations are chosen at random for
P5
further study. For reasons of comparison and processing, the
simulations are reblocked to 15 × 30 × 7.5 m3 blocks, which is the
2200 block size used at the mine. Grade tonnage curves are shown in
South Figure 4. It is apparent from the figure that for the model based on
P1
exploration data, the average grade above cutoff is higher at all
P2 cutoffs. The amount of tonnes above cutoff is lower compared to
the updated models.
P3
Grade tonnage curves
P4 12
Millions

100.0
Y=99599 10
2000
80.0
8

Grade g/t
Tonnage

50200 50400 60.0

6
Period (years)
40.0
4
1 2005
20.0 2
2 2006

3 2007 0.0 0
0 0.4 0.8 1.2 1.6 2 2.4 2.8 3.2 3.6 4 4.4 4.8
4 2008 Cut-off grade g/t

5 2009
FIG 4 - Grade tonnage curves for the models based on exploration
data (dotted lines) and the models updated with simulated future
grade control data (solid lines).
FIG 3 - East-West sections of the yearly schedule in a gold mine.

Figure 5 displays the ore tonnes and metal recovered when

in 2004 was about 43 Mt, which drops to about 26 Mt in 2005
and again to 11 Mt in 2007. This will necessitate a resizing of the
equipment fleet sometime in mid-2005. Figure 3 depicts the
long-term (yearly) production schedule in use at the mine. The
mine’s schedule proceeds one layer at a time and indicates that
the pit will be exhausted by the end of the fifth year. At this
mine, blasting occurs on 7.5 m benches and grade control is done
using reverse circulation drilling on 5 m × 7 m spacing. The plant
capacity is 3.6 Mt/a and the mill circuit consists of crushing and
grinding after which the gold is recovered through a Carbon in
Leach (CIL) circuit.
mining is done according to the mine’s yearly schedule (as
shown in Figure 3) when short scale deposit information is taken
into account. The risk profiles were generated by running the 20
updated models through the mine’s yearly schedule. During this
process the tonnes of ore/metal produced and NPV generated by
each updated model was recorded and plotted on Figure 6. The
updated models indicated that the amount of ore tonnes produced
is not enough to fill the plant to capacity (3.6 Mt/a) for all the
years from 2005 to 2009. They also indicate a 50 per cent chance
of negative NPV in 2005 when insufficient metal is produced.

324 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 STOCHASTIC MINE PLANNING – EXAMPLE AND VALUE FROM INTEGRATING LONG- AND SHORT-TERM MINE PLANNING

Risk profile for ore production for Sunrise Dam long-term schedules derived using orebody models based on two sets of
schedule; updated models information. As a result, it provides a look at the consequences of
5.0 not taking closer spaced grade control information into account
Millions

4.5 when performing production scheduling.

4.0 The production schedules were generated for four month
3.5 periods (short-term schedules). The production capacity for each
four month period was set at 1.2 MT and the targeted grade
Tonnage

3.0
2.5 constraints between two and 7 g/t. To make the schedules practical
2.0 for equipment mobility and space, the schedules were smoothed.
1.5 After smoothing, the simulated orebody models are run through
1.0 their respective schedules to generate risk profiles for ore, metal
0.5 and NPV. Each short-term schedule was regrouped to generate the
0.0 yearly LOM schedule of the mine. The risk profiles of the
2004 2005 2006 2007 2008 2009 2010 production schedule derived from the 20 exploration based models
Years for ore, metal and NPV are shown in Figure 6. As the figure
Risk profile for ore production
Risk profile of metal content for Sunrise Dam long-term schedule; Schedule derived from simulations based on exploration data
updated models 4.3
Millions grams

Millions
18
4.1
16
3.9
14 3.7

Tonnage
12 3.5
Metal content

10 3.3

8 3.1

6 2.9

4 2.7

2
2.5
2004 2005 2006 2007 2008 2009
0 Periods (years)
2004 2005 2006 2007 2008 2009 2010
Years
Millions grams

Risk profile for metal production

Schedule derived from simulations based on exploration data
Risk profile of NPV for Sunrise Dam long-term schedule; 22
updated models 20
18
250.0
Millions AUD

16
Metal content

14
200.0
12
10
150.0 8
6
NPV

100.0 4
2
50.0 0
2004 2005 2006 2007 2008 2009
Periods (years)
0.0
2004 2005 2006 2007 2008 2009 2010 Risk profile for NPV
-50.0 Schedule derived from simulations based on exploration data
Years 250.0
Millions AUD

200.0
Mill target (ore production)
Average of the simulations 150.0
NPV

Simulations 100.0

FIG 5 - Performance of the mine’s yearly schedule when local 50.0

orebody information is taken into account.
0.0
2004 2005 2006 2007 2008 2009

Deriving production schedules using the SIP Periods (years)

formulation
The following section describes the process, as given in Figure 1, Mill target (ore production)
used to derive the production schedules. First the 20 simulated Average of the simulations
orebody models conditioned on exploration data are put through
the SIP formulation in Equation 1. Next, the 20 updated orebody Simulations
models are then run through the SIP formulation and a production
schedule accounting for short-scale orebody information is FIG 6 - Risk profile of ore, metal and NPV for the production
derived. This methodology allows for a comparison of the schedules derived from orebody models based on exploration data.

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 325
A JEWBALI and R DIMITRAKOPOULOS

shows, the yearly production target of 3.6 Mt/a is met for 2005 to
2007. Note that this schedule predicts that the pit will be
exhausted by 2008. Figure 7 describes the schedule derived from
the updated models. This schedule extends the life of mine by
another year to 2009. As the figure shows, the short-term schedule
has no problem delivering 3.6 Mt/a of ore in all years except in
2009, when the last of the ore will be mined.
The schedules based on updated models produce 3.6 Mt more
ore and as a consequence 2.6 million grams more metal, which
results in a cumulative NPV that is A$7.7 M higher. The
production schedules based on orebody models derived from
Risk profile for ore production
Schedule derived from updated models
4.7
Millions
exploration data (that do not account for local orebody
knowledge and characteristics) underestimate the actual tonnes
of ore mined from the orebody. The derived exploration based
schedules are a conservative option. Taking local grade control
information into account indicates that there is probably
sufficient ore to extend the life of mine to 2009 compared to
2008 when only exploration data is used. Figure 8 displays the
production schedules for the exploration based simulations and
the updated simulations. As Figure 8 shows, the two schedules
are different in that different sections of the pit are mined out in
different time periods. However, when compared to Figure 3,
both of the SIP derived schedules mine the pit out in a series of
successive cones, while the mine’s yearly schedule proceeds in
layers.

4.5
4.3
TABLE 1
4.1 Cumulative averages for ore tonnes, metal and NPV over all
3.9 simulated orebody models (both exploration data based and
Tonnage

3.7 updated models).

3.5
3.3
SIP and SIP and Mines schedule
simulations updated performance
3.1
with simulations with updated
2.9
exploration with ‘future’ simulations
2.7
data data
2.5
2004 2005 2006 2007 2008 2009 2010 Ore tonnes 14 18 10
Periods (years) (million grams)
Metal tonnes 52 55 38
Risk profile for metal production (million grams)
Schedule derived from updated orebody models
18 NPV 552 560 330
Million grams

(million AUD)
16

14

12
Metal content

10 2200
P4 P1
8
P2 P1 P3
6
P5 P4
4
2000 Y=100089
North
2

0
2004 2005 2006 2007 2008 2009 2010 P5
2200
Periods (years)
P1
P1
P2
Risk profile for NPV P3
Schedule derived from updated models P2
2000 Y=99824 P3
Millions AUD

200.0
P4
P4

150.0
2200
P3 P3
P1
NPV

P5 P2
100.0
Y=99599 P4

2000 South
50.0
50200 50400 50200 50400

0.0 Period (years)

2004 2005 2006 2007 2008 2009 2010
Periods (years)
1 2005
2 2006
3 2007
Mill target (ore production)
4 2008
Average of the simulations
5 2009
Simulations

FIG 7 - Risk profiles of ore, metal and NPV for the production FIG 8 - West-east sections of the Sunrise pit displaying production
schedule derived from orebody models displaying grade control schedules derived from simulation based on exploration data
characteristics. (right) and updated models (left).

326 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I
 STOCHASTIC MINE PLANNING – EXAMPLE AND VALUE FROM INTEGRATING LONG- AND SHORT-TERM MINE PLANNING
CONCLUSIONS
A new multistage stochastic mine production scheduling
approach presented herein includes four stages:
1. Stage 1: generates high density future grade control data for
incorporation into the production scheduling process. It is
based on sequential Gaussian co-simulation of exploration
and grade control data from previously mined out parts of a
deposit.
2. Stage 2: employs conditional simulation by successive
residuals to update pre-existing simulated orebody models
with the simulated future grade control information from
Stage 1.
3. Stage 3: uses a stochastic integer programming mine
scheduling formulation and balances both maximising net
present value and minimising deviations from production
targets.
4. Stage 4: includes quantification of risk in the produced
schedules generated and reporting.
The application at a large operating gold mine demonstrates
that the proposed approach is practical and adds value to the
operation. The approach is shown to deliver 3.6 Mt of additional
ore and 2.6 million grams more metal – which matches better
with the mine’s reconciliations. The approach also results in a
cumulative NPV which is on average $7.7 M higher that that of a
comparable stochastic schedule without the simulated grade
control data and substantially higher when compared to the NPV
from the actual schedule of the mine.
ACKNOWLEDGEMENT
The authors are grateful to AngloGold Ashanti for funding and
multifaceted support, Rio Tinto for funding, as well as additional
support from NSERC grants and McGill’s COSMO Stochastic
Mine Planning Laboratory.
REFERENCES
Benndorf, J and Dimitrakopoulos, R, 2010. Stochastic long-term
production scheduling of iron ore deposits – Integrating joint
multi-element geological uncertainty, in Advances in Orebody and
Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 151-158 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Dimitrakopoulos, R and Abdel Sabour, S A, 2007. Evaluating mine plans
under uncertainty: Can the real options make a difference? Resources
Policy, 32:116-125.
Dimitrakopoulos, R, Martinez, L S and Ramazan, S, 2007. A maximum
upside/minimum downside approach to the traditional optimisation
of open pit mine design, Journal of Mining Science, 43:73-82.
Advances in Orebody Modelling and Strategic Mine Planning I
Godoy, M and Dimitrakopoulos, R, 2004. Managing risk and waste
mining in long-term production scheduling of open pit mines:
Transactions of Society for Mining, Metallurgy and Exploration,
316:43-50.
Goovaerts, P, 1997. Geostatistics for Natural Resources Evaluation, 483 p
(Oxford University Press: New York).
Guardiano, E, Parker, H M and Isaaks, E H, 1995. Prediction of
recoverable reserves using conditional simulation: A case study for
the Fort Knox gold project, Alaska, Unpublished report, Mineral
Resources Development Inc, Foster City.
Journel, A G and Kyriakidis, P C, 2004. Evaluation of Mineral Reserves:
A Simulation Approach, 216 p (Oxford University Press: New York).
Khosrowshahi, S, Shaw, W J and Yeates, G A, 2007. Quantification of
risk using simulation of the chain of mining – A case study on
Escondida Copper, in Orebody Modelling and Strategic Mine
Planning, second edition (ed: R Dimitrakopoulos), pp 33-42 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Leite, A and Dimitrakopoulos, R, 2007. A stochastic optimisation model
for open pit mine planning: Application and risk analysis at a copper
deposit, Transactions of the Institutions of Mining and Metallurgy,
Mining Technology, 116:A109-A118.
Leite, A and Dimitrakopoulos, R, 2010. Production scheduling under
metal uncertainty – Application of stochastic mathematical
programming at an open pit copper mine and comparison to
conventional scheduling, in Advances in Orebody Modelling and
Strategic Mine Planning I (ed: R Dimitrakopoulos), pp 35-40 (The
Australasian Institute of Mining and Metallurgy: Melbourne).
Meagher, C, Abdel Sabour, S A and Dimitrakopoulos, R, 2010. Pushback
design of open pit mines under geological and market uncertainties,
in Advances in Orebody Modelling and Strategic Mine Planning I
(ed: R Dimitrakopoulos), pp 291-298 (The Australasian Institute of
Mining and Metallurgy: Melbourne).
Myers, D, 1991. Pseudo cross-variograms, positive definitness and
cokriging, Mathematical Geology, 14:805-816.
Peatie, R and Dimitrakopoulos, R, 2009. Forecasting recoverable ore
reserves and their uncertainty at Morila gold deposit, Mali – An
efficient simulation approach and future grade control drilling, in
Proceedings Orebody Modelling and Strategic Mine Planning (ed: R
Dimitrakopoulos), pp 193-200 (The Australasian Institute of Mining
and Metallurgy: Melbourne).
Ramazan, S and Dimitrakopoulos, R, 2007. Stochastic optimisation of
long-term production scheduling for open pit mines with a new
integer programming formulation, in Orebody Modelling and
Strategic Mine Planning, second edition (ed: R Dimitrakopoulos),
pp 385-392 (The Australasian Institute of Mining and Metallurgy:
Melbourne).
Ramazan, S and Dimitrakopoulos, R, in press. Production scheduling
with uncertain supply – A new solution to the open pit mining
problem, European J Operations Research and COSMO Lab
research report no 1, 2007, pp 257-294.
Vargas-Guzman, J A and Dimitrakopoulos, R, 2002. Conditional
simulation of random fields by successive residuals, Mathematical
Geology, 34:507-611.
Spectrum Series Volume 17 327
 HOME

Author Index
A E L
Abdel Sabour, S A 29, 291 Elkington, T 241 Leenders, B 113
Acuña, E 263 Leite, A 35
Allen, L 193, 199 Lopes, J A 205
F
Arvidson, H 43
Fava, L 263
M
Fereira, J 79
B Fernandes, J B 205 Malajczuk, W 101
Bandarian, E M 79 Filippov, P A 129 Meagher, C 291
Benndorf, J 151, 167 Flitton, T 257 Mueller, U A 79
Bodon, P 277 Freidin, A 249 Mustapha, H 309
Botelho, M C 123 Fricke, C 277 Myers, P 241
Boucher, A 59
Boyle, C 67
N
Bray, A 223
G
Neverov, A A 129
Brazil, M 235 Gloaguen, E 91
Neverov, S A 129
Brisebois, K 299 Godoy, M 21, 215
Groeneveld, B 113
O
C
Oparin, V 249
Carter, P 271 H
Osterholt, V 43
Chitiyo, G 223 Hall, S 263
Chou, C L 215 Hammitt, J 299
Coutts, B 143 Hardcastle, S 263
P
Hardtke, W 193, 199 Peattie, R 257
Hart, W 299 Perry, B 193, 199
D
Herod, O 43
de Carvalho Pereira, G J 123
Hilário, J A 123 R
Dimitrakopoulos, R 13, 29, 35
91, 151
Rendu, J-M 173
215, 291
309, 321 J Reynolds, K 159
Dincer, T 159 Jewbali, A 321 Richardson, S 79
Douglas, I 193, 199 Richmond, A 285
Dowd, P A 51 Robins, S 179
Du Pisani, P 223 K Rocchi, L 271
Dunn, P 263 Kemp, C 223 Rodriguez, P C 123
Durham, R 241 King, B 7 Rosas, C F 205

Advances in Orebody Modelling and Strategic Mine Planning I Spectrum Series Volume 17 329
S Topal, E 113 X
Sandeman, T 277 Xu, C 51
Silva, E A 123 V
Stanford, C 277 van der Riet, J 101
Z
Stone, P 101, 271 Vanzela, G A 205
Zuckerberg, M 101
Verly, G 299
Volz, M G 235
T
Tapsiev, A 249
Thomas, D A 235 W
Tomaselli, B Y 123 Wharton, C 107
Tonkin, C 135 Whittle, J 3

330 Spectrum Series Volume 17 Advances in Orebody Modelling and Strategic Mine Planning I