Professional Documents
Culture Documents
Preprint 04-120
A. Proulx
INCO, Limited
Thompson, MB, Canada
O. Tavchandjian
INCO Technological Services Ltd.
Sudbury, ON, Canada
A. Aubut
INCO Technological Services Ltd.
Thompson, MB, Canada
Mineral envelopes were based on host lithologies and the sub- length approximates the routine sample length for significant miner-
sequent structural deformation, which control the emplacement of the alization.
ore mineralization. Drilling information was supplemented with cross- The mineralization envelopes vary along strike and dip with
hole seismic tomography surveys, oriented core measurements, and pinch and swell geometries. Cartesian distances do not reflect the
in-hole Gyro location surveys. These surveys helped in establishing natural geological distances between samples. Therefore, an
the shapes and extents of the envelopes. unfolding procedure (DATAMINE, 1997) was applied to ensure
The 1D Lower ore body contains at least three different zones appropriate geological distances between samples for the geostatis-
of mineralization: Footwall zone (FW), Mid zone (MID), and Hanging tical modeling. This unfolding procedure resulted in the creation of a
wall zone (HW) (Figure 1). The FW zone mineral envelope was divid- new set of co-ordinates (UCSA [across direction], UCSB [down dip
ed into two mineral domains: Fw1 and Fw2. The MID zone was sub direction], and UCSC [along strike direction]) for each sample and
divided into three mineral domains: Mid, S_mid and S_fw. The each discretization point of the deposit.
Hanging wall zone contained a single domain (Hw) (Figure 1). The To properly weight the grade by density, a service variable was
domain envelopes were created loosely by allowing for 1.52 m (5 ft) used for the construction of both the interpolated and the simulated
of waste along their margins. This approach ensured that the grade models. This service variable is defined as QNi = Ni x SG. The aver-
estimates would not be biased by an overly constrained interpreta- age percentage Ni of each block was back calculated after modeling
tion. Domain overlapping was used to capture samples, which could by dividing the estimated QNi by SG
potentially belong to more than one domain.
Variography
Project Database
The experimental variograms of QNi were calculated for the
Approximately 700 drill-holes outline the assessment area. The composite sample sets using the unfolded co-ordinates. The
drill-hole collars are located approximately 15 m (50 ft) apart on East- anisotropy was evaluated by analysis of variogram fans every 13
West oriented sections with some minor off-section drilling. degrees in the strike-dip plane of the mineralization.
Core recovery is excellent. Implicitly missing assays were The QNi grade variograms were then modeled by the linear
assigned background default values while explicitly missing assays combination of a nugget effect and two spherical structures (Figure
were replaced by nil values. 2).
The Specific Gravity was calculated using a multi linear rela- As expected, the grade variograms were found to be signifi-
tionship based on Ni, Fe, S, and Cu grades: cantly different in each domain. This result confirmed the justification
SG = 100 / (100 / 2.6633 - 0.1833 x Cu - 0.412 x Ni - for the separate modeling of the domains.
0.3969 x Fe + 0.2114 * S) (Eq. 1)
Declustered Statistics
Where Declustered statistics were computed within the domains using
SG= specific gravity in grams per cubic centimeter a nearest neighbor (NN) block model. The NN model was built using
Cu = assayed copper by percentage an anisotropic search ellipsoid, defined from the longest range of the
Ni = assayed nickel by percentage QNi variogram. This model included the transformation of the
Fe = assayed iron by percentage Cartesian distances into geological distances using the unfolding
S = assayed sulfur by percentage algorithm of the DATAMINE software.
The QNi declustered statistics are presented in Table I for the
individual domains. All the elements showed skewed frequency dis-
All the drill-hole assays were captured in each respective tribution reflecting the dominance of low-grade mineralization with a
domain and stored in separate files. All assays were composited on few higher values related to massive sulfide mineralization.
a uniform length of 1.52 m (5 ft) weighted by SG. The composite
HW EXPLOR
ATION DR
IFT
3500 LEVEL
FW ZONE
HW ZONE
N
FW RAMP
MID ZONE
3D VIEW
Kriging Interpolation
Interpolation techniques based on kriging algorithms aim at min-
imizing the variance of the error on the estimate of the local average
grade (Matheron, 1963; Journel and Hjuibregts, 1983). The metal
grade is usually estimated for small blocks corresponding to the
smallest mining unit (SMU) that can be considered to discriminate
between ore and waste for mine planning.
In the present study, Multiple Indicator Kriging (MIK) was select-
ed as an appropriate interpolation method in order to deal with the
skewness in the grade frequency distribution (Journel, 1983). MIK is
based on binary indicator coding of grade data. Each grade field is
replaced by a binary function; coding the presence or the absence of
a metal grade above a given threshold. A local conditional distribution
can be built for each block from the kriged estimates of the indicator
Mid domain functions, from which an average grade can be estimated.
In each domain, all of the composite samples were coded
according to QNi indicator cut-offs. These were determined from the
declustered cumulative statistics and each represents approximately
9% of the total metal in the domain. Indicator variograms were mod-
eled by a linear combination of a nugget effect and two spherical
structures. In order to compare the various indicator variograms, they
were normalized to their respective variance. As expected, high-
grade indicators show a shorter range (i.e. lower spatial correlation)
than low-grade indicators
Each block in the block model was discretised by a matrix of 2
x 3 x 3 points. The indicator functions were interpolated at each
unfolded discretization point using the search volume ratios defined
by the second range of the median indicator variogram (within the
unfolded plane of the mineralization) A nested sample search was
used. It applied octant constraints on sample selection. For blocks
not estimated during this first search, expanded search ellipsoids,
with reduced constraint restrictions on the minimum number of sam-
Hw domain ples and octants, were applied. Blocks estimated during the least
constrained searches were flagged in the model. The estimates in
these blocks may be locally imprecise and have a potentially large
conditional bias.
Fw domain
Block Model Validation Table II Statistical description of QNi in the MIK model
The first check on the validity of the block model was a compar-
ison between the block grade spatial distributions with the cross sec-
tional drill-hole data. The unfolding procedure significantly improved
the spatial distribution of the grade (Figure.3). Trends in the block
grade spatial distributions honor the geological trends. The statistics
on QNi derived from the MIK model are presented in Table II. These
can be compared to NN model statistics for each domain in Table I.
As expected, the MIK models have a similar global mean to the
NN model estimates (Table I) but have a significantly lower variance
and a reduced skewness. The lower variance and skewness are
expected since the NN model represents the sample population and
not the block population.
Volume-Variance Adjustment
Gamma(h)
algorithm of the DATAMINE software).
1
Gaussian Transform 0.8
A requirement of any Gaussian algorithm is an assumption of 0.6
normality. Since the original sample grades were not normally dis- 0.4
tributed, the declustered histograms of QNi, obtained using the NN
0.2
interpolation (Table I) were used to build graphical anamorphoses of
0
QNi values to normalized Gaussian values with a mean of 0 and a
0 50 100 150 200
variance of 1. The theoretical model statistics were obtained by prop-
erly assigning each Gaussian value, its declustering weight. Since h (feet)
the model considered here is an anamorphosed Gaussian random Variogram Theoretical Experimental
function, it is important to test the compatibility of the data with the
selected model. A number of tests can be used for an anamorphosed
Figure 4 Bigaussian Normality check on Hw domain sam-
Gaussian random function. A check used by many authors (Verly,
ple data
1983) is to calculate the ratio between the square root of the vari-
ogram (?2) and the madogram (1). This ratio should present a con-
stant value of 1.77 (i.e., ? ). This check was carried out on the raw
data for each domain. Although these checks cannot be considered Gaussian Variograms
as totally conclusive, (they did not take into account the declustering The experimental variograms of the Gaussian transformed val-
weights and the number of pair of points was fairly restricted) they do ues were calculated in the unfolded coordinate system: UCSA, UCSB
not indicate any significant problem with the proposed models. Figure and UCSC. Modeling of the variograms was carried out for each
4 illustrates an example of the graphical results of a bigaussian nor- domain (Table III).
mality check.
Table III Summary of variogram modeling parameters for the Gaussian transformed values of QNi
To ensure that the simulated variograms corresponded well to of all the simulated points was also produced and compared against
the model variogram, the simulations were randomly sampled and the input histogram. This initial check verified that the non-condition-
variograms were calculated from the sampled points. The histogram al simulations were performing well.
Conditioning of Simulations Using Simple Kriging For all the conditional simulations, the final global mean and
variance for QNI, Ni and SG, were consistent with the expected val-
The non-conditional simulations were forced to honor the actu- ues defined by the declustered statistics and variograms.
al data using the following procedure:
For all the nodes located within a distance corresponding to CALIBRATION AND VALIDATION OF SIMULATIONS IN
approximately 90% of the total sill of the Gaussian variogram, calcu- THE UPPER 1-D ORE BODY
late the conditioned simulated value (Ycs) as:
Ycs = Ysk + (Y1TB Y1sk) (Eg.2) In order to validate the application of the results of the condi-
Where: tional simulations to the assessment of recoverable resources and
Ysk is the Kriged value using actual samples reserves, a calibration exercise was undertaken. A mined-out portion
Y1TB is the turning band simulation at node location of the ore body was selected as a back-analysis case study. This
Y1sk is kriged value using simulated samples area is similar geologically to the advanced exploration portion of the
For all the other nodes set Ycs to Y1TB since simple kriging of ore body. The case study area was well-drilled (drill-holes were col-
N(0,1) would potentially suffer from significant conditional bias and lared on 7.6 m [25-foot] and 15 m [50-foot] centers) and has exten-
result in nil very low values for both Ysk and Y1sk. sive mapping of underground openings. The record of the mill-credit-
At this stage, a second validation check was performed to ed production was also available. However, the credited grade rep-
ensure that the conditional simulations still respected the input model resents a prorated credit back to the grade reported as mined based
variogram and histogram. on production tonnage and grade estimates.
Back Transformation of Simulated Gaussian Values to QNi The methodology applied was as follows:
and Reconstitution to Ni Values Create an exploration data set by removing all in-fill produc-
Conditional simulations of Gaussian values were back trans- tion drilling.
formed to QNi using the inverse graphic anamorphosis based on the Create mineral envelops from the exploration data set (15 m
cumulative histograms. Histograms of the back transformed data and 30 m [50 and 100-foot] drill spacing).
were compared to the original declustered histograms of QNi for each Perform NN, MIK, and CSs modeling from the exploration
mineral domain. The agreement was generally very good (Figure 5). data set.
Assess the recoverable resource at various cut-off grades
based on the exploration data set.
Create mineral envelops from in-fill drilling data set. (7.6 m
[25-foot] drill spacing).
Perform MIK and CSs modeling using all in-fill drilling from
the production data set.
Assess the recoverable resource at various cut-off grades
based on the in-fill drilling data set.
Compare historical credited production, NN, MIK and CSs
results based on both, the exploration dataset, and the pro-
duction data set (based on bulk-mining only).
Compare the actual detailed ore mapping from mine open-
ings to the spatial patterns produced by NN, MIK and CSs
models.
Although a large number of publications present successful
applications of CSs, few are for underground mining (Ravenscroft,
1993; Davis, 1999). This is deemed related to the additional difficulty
of handling the large number of simulations when performing mine
Declustered Simulated
QNI QNI
Finally, the SG and Ni values were back calculated from the QNi
values using the following linear relationship based on the declus- NC SIMULATION3 MIK
3200L MAPPED ORE NN
tered statistics:
SG = 0.03 x QNi + 2.82 (Eq. 3)
Ni = QNi/SG (Eq. 4)
Figure 6 Comparison of actual ore mapping, CS, MIK and NN
models in the calibration zone
1.0
3.00
Recoverable Resource Tonnage
2.50
0.6
2.00
0.4
1.50
0.2
1.00
1.00
1.25
1.50
1.75
2.00
0.0
1.00
1.25
1.50
1.75
2.00
%Ni Cut-off
% Ni Cut-off NN_EXPL IK_DEFN
NN_EXPL IK_DEFN MILL CREDIT IK_EXPL
MILL CREDIT IK_EXPL CS 6 CS 7
CS 6 CS 7 90% LCL 90% UCL
90% LCL 90% UCL
Figure 7 Comparison of tonnage and grade curves- obtained from the various assessment methods and mill credit in the calibra-
tion zone (CSs 6 and 7 represent the range of 90% of simulated outcomes, while the 90% control limits represent a +/-10% differ-
ence from the median simulation)
planning for underground deposits, which is more complex than in USE OF CS TO COMPARE TWO MINING METHODS FOR
most open pit situations. The FSO was used to quickly assess the THE 1-D LOWER ORE BODY
recoverable resource from all the block models and simulations at
various cut-off grades. This is similar to the floating cone algorithm Based on the results obtained during the validation exercise, CS
used for open pit situations. It provides a close to finish product, was selected as a better tool to allow mine planners to compare bulk-
which may be used to effectively compare alternative mining scenar- mining and cut-and-fill mining in the 1-D Lower ore body. Twenty CSs
ios for conceptual planning exercises. were created for all six mineral domains. The mineral domains and
Results indicated that the recoverable tonnage and grade are samples used in the MIK block modeling were used to constrain and
better predicted using the CSs than the NN or MIK block models condition the simulations.
(Figure 7) when the exploration data set is used. As well, the spatial A randomly selected simulation was chosen as a base-case.
patterns created by the CSs, more closely resembles those defined Detailed mine planning for the two mining options was performed on
by the detailed ore mapping of the existing openings (Figure 6). The this simulation model. Comparison of the FSO runs to manual plan-
results show that when little information is available for estimation, ning performed on the simulation, verified the FSO parameters.
the spatial continuity of the high grade and low grade is over-stated These results are illustrated in Table IV and Figures 8a and 8b, for the
in the NN and MIK block models. This produces an over-estimation bulk-mining scenario. The detailed mine planning also allowed for the
of the recoverable grade and an under-estimation of the recoverable development of global factors for mining recoveries and dilution. All
tonnage for the NN model (NN_EXPL) and, a general overestimation subsequent FSO runs on the simulation models and estimation mod-
of both tonnage and grade for the MIK model (IK_EXPL), when com- els used these same parameters.
pared to the historical mill-credited production and the MIK model Figures 9a and 9b summarize the results of this comparative
(IK_DEFN) based on the production data set (Figure 7). study in the form of grade and tonnage curves. The results obtained
with the NN and the MIK models are also plotted on these graphs for
reference. These series of curves show that the MIK model provides FSO to the NN model produces erroneous results, except for the
a similar answer to the CSs for the bulk-mining scenario. However, recoverable grade in the cut-and-fill scenario. The most likely expla-
there is a significant difference between the MIK model and the CSs nation for this result is that the dilution included by the FSO counter
in the cut-and-fill scenario. This result implies that making a capital balanced the high grading effect of the NN model. In addition, these
investment decision based on only a MIK model, would present a sig- graphs show that the range of simulations provides a tool to perform
nificant risk of not selecting the best mining method. Applying the meaningful risk and sensitivity analysis on each option.
Table IV Summary of FSO parametersbulk-mining option as compared to manual planning on a base-case simulation
Figure 8a Plan view of manual bulk-mining stopes (white Figure 8b Plan view of FSO Run1 bulk-mining
lines) stopes (white lines)
12.5
12.5
10.5
10.5
Recoverable Tonnage
Recoverable Tonnage
8.5
8.5
6.5
6.5
4.5
4.5
2.5
2.5
1.00
1.25
1.50
1.75
2.00
1.00
1.25
1.50
1.75
2.00
%Ni Cut-off %Ni Cut-off
CS 19 CS 20 MIK CS 19 CS 20 MIK
NN 90% LCL 90% UCL NN 90% LCL 90% UCL
3.10
3.50
2.90
2.70
Recoverable Resource Grade
3.00
2.50
2.30
2.50
2.10
1.90
2.00
1.70
1.50 1.50
1.00
1.25
1.50
1.75
2.00
1.00
1.25
1.50
1.75
2.00
% Ni Cut-off % Ni Cut-off
CS 19 CS 20 MIK CS 19 CS 20 MIK
Figure 9a Comparison of bulk-mining grade and tonnage Figure 9b Comparison of the cut-and-fill mining grade
curves- obtained with the various assessment methods (CSs and tonnage curves- obtained with the various assess-
19 and 20 represent the range of 90% of simulated outcomes, ment methods (CSs 19 and 20 represent the range of 90%
while the 90% control limits represent a +/-10% difference of simulated outcomes, while the 90% control limits rep-
from the median simulation) resent a +/-10% difference from the median simulation)