7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

DOI: 10.7251/BMC170701189E

IMPLICIT 3D SOLID MODELLING WITH SUPPORT MACHINES

G. ERTUNÇ1, F. ATALAY1
1
Department of Mining Engineering, Hacettepe University, Beytepe, 06800, Ankara, TURKEY
gertunc@hacettepe.edu.tr, atalay@hacettepe.edu.tr

ABSTRACT

Solid modelling in mineral resource estimation is important phase since block model is built with
respected to the orebody. If the solid model of the ore deposit is not consistent with approximate real
case, error will accumulate progressively through the estimation steps. Interpretation in classical
approach, where sections are taken through drillholes and polygonized with geological and mining
knowledge, is very critical and the process is always time consuming. In this study, support vector
machine (SVM) algorithm compiled by authors is offered as a solution for 3D solid modelling. The
proposed implicit method is based on Gaussian radial basis functions and set of northing sections are
classified as ore zone and no-ore zone, respectively. The input variables are indicator values derived
from boreholes of correspondant sections and the output, 3D solid, is created by combining all classified
sections. The SVM is fast when compared to the classical slice sectioning and 3D solid model reveals
consistent result.

Key words: Orebody, Support vector machines, 3D solid model.

1. INTRODUCTION

It is an undeniable fact that mining plays an vital role in the development of countries through
the rational use of science and technology. When countries with a high added value based on
mining are examined, it appears that almost all of the mining activities are done within the
framework of a resource-reserve coding standards. These codes are listed such as
NI43-101/CIM [1], JORC [2], PERC [3], SAMREC [4], which are collected under the umbrella
of the CRIRSCO (Combined Reserves International Reporting Standards Committee). In
Turkey, nowadays content and the framework of the repoting standard UMREK (National
Resources and Reserves Reporting Comittee) commenced.
Report standards are classified according to a certain level of confidence in the presence of
underlying orebody. Although the level of confidence is defined on the basis of geostatistical
estimates or simulations at the centroids of the block model, it is directly dependent on how
closely the orebody is modeled with respect to the ground-truth.For example, a geological
section through deposit where no ore is present will lead to the formation of a block pattern, in
other words a non-existent ore presence. Thus, more precise ore-bed rock discrimination is
achieved, the less systematical error will appear through the steps of resource estimations.

189
 7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

Today, with the development of mining software, the sections taken by hand are digitally
transferred to the computer environment and offer the user instant 3-dimensional visualization.
Each section drawn by hand is actually an interpolation. The user creates a polygon by hand,
taking into account the impact distance of the borehole on the section taken, and a series of
polygons then forms the volume of orebody. Every error made in this sectioning method is
systematically carried to every stage extending to the end of the mineral resource estimate.
Since each user's geologic interpretation will be different from each other, there is no criterion
of how much error is made in the sectioning process.
Orebodies are formed by the combination of various geological and atmospheric laws and
appropriate conditions in the earth's crust. Naturally, the orebody is systematically
differentiated from other rocks and is located in a certain pattern. Support vector machines
method, which has an important place among the machine learning methods, has been used to
reveal this pattern in the study. Through this method, the data patterns arising from the analysis
of the existing data are categorized and classified. In this respect, the grades below a certain
threshold and not to be used in mine planning are taken into consideration during the sectioning
process. An algorithm that classifies the boreholes based on cross-sections according to their
threshold grade is compiled with the MATLAB program. The proposed implicit method is
based on Gaussian radial basis functions and it is applied to an epithermal gold deposit in order
to build 3D solid model of the mineralization. Mineralized zones in each section is classified
with SVM based on indicator variables of drill-holes falling in corresponding sections and
finally these sections are combined in order to create 3D solid model. In addition, generalized
accuracy of the classification is established by cross validation by test and validation data.

2. METHODOLOGY

Support vector machines are set of related supervised learning methods used for classification
and regression and they are regarded as generalized linear classifiers. SVMis based on the
quadratic optimization problem, and does not require any knowledge of the distribution
function for the data. In SVM classification, data optimally separated into the two categories of
n-dimensional hyperplane (hyperplane) form. In the case of binary classification, the aim is to
find decision boundary that partitions the underlying vector space into two sets, one for each
class. The formula for the output of a linear SVM is given as:
( )
where w is the normal vector to the hyperplane, x is the input vector, and b is the margin offset.
The separating hyperplane is the plane ( ) . The nearest points lie on the planes
( ) .
The margin is defined by the distance of the hyperplane to the nearest of the positive and
negative examples. The aim of the SVM is to obtain maximum margin that is
subjected to ( ) where where is the th training example, and is the
correct output of the SVM for the th training example. The value is +1 for the positive
examples in a class and –1 for the negative examples.
Finding maximum margin is defined as an quadtratic program (QP) problem in which the
objective function is dependent on Lagrangian multipliers namely, . The dual formulation
yields:

190
 7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

∑∑ ( ) ∑

under constraints, and ∑

where N is the number of training samples,

Cortes and Vapnik [5] introduced a modification to the original optimization statement which
allows the failure of an example to reach the correct margin. But it is penalized with slack
variable , wide margin trade-off parameter C. Then, the dual formulation is modified under
constraint .

Non-linear SVM classifiers are generalized as follows:

( ) ∑ ( )

where ( ) is kernel function between input vector and training vector .

The nonlinear case slightly alters the quadratic form but are still computed by QP such as,

∑∑ ( ) ∑

with constraints and∑

There are numerous kernel functions available for SVM, such as Gaussian, polynomials, and
neural networks. In this study, Gaussian Radial Basis Function Kernel (GrbfK) ( )
‖ ‖
is applied.

The QP problem is solved when three conditions are met for all , , , and
.

There are many different kernel functions that DVM uses in nonlinear distinction situations.
Frequently used functions are Gauss, polynomials and neural networks.Quadratic problem
programs (QP) are often processes that result in complex equations and long solution times.
Sequential Minimal Optimization (SMO) method which partitions large QP problems into
small sections is used in the study and detail information can be referred in Platt [6].
The method of binary classification is adhered to several classifications by some modifications.
Two main methods for multiple classification in the literature have been proposed. These are
One-vs-all (OVA) [7] and One-vs-one (OVO) [8].

191
 7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

3. CASE STUDY

In the field, 22 sections are created with bandwidth of 50 m and each section is classified by
means of high grade ore, mid-grade ore and waste (red, blue, and black, respectively) (Figure
1). Soil samples and assay values from boreholes which fall into corresponding bandwidth are
used as an input in each slice and total data count is 23150. Results of the each slice then
imported into block model centroids in order to have better visualization of the orebody (Figure
‖ ‖
2). In this study, Gaussian kernel function is used which is defined by ( ) .
10-fold cross-validation was applied for training and validation steps.In the SVM algorithm, the
Gaussian kernel function parameter pair (C, γ) automatically selected for the generalized
accuracy. The parameter pair which gives the highest accuracy value according to the learning
and test data is used as the prediction parameter. In the case of more than one parameter pair
giving equal accuracy value, the pair that gets faster solution time is selected.

Figure 1 Sections, soil, and borehole data.

192
 7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

Figure 2 Orebody geometry generated from

The blocks categorized according to the threshold grade output by the algorithm are given in
Figure 3. When the results are evaluated, it is seen that the results are consistent with the
borehole data and threshold values are also honored.
Blocks estimated using SVM are indicative for the geological section and can be used as
templates at the final stage of the final determination of the ore boundaries. This template is
transferred to the mining software and the section is digitized by hand drawing in this
environment. The final result is shown in Figure 4.

Figure 3 One example support vector machine section. Orebody section (blue: low grade, red: high grade)

193
 7th BALKANMINE CONGRESS – Proceedings, Prijedor 2017

Figure 4Interpreted digitized section of an example section

4. CONCLUSION

In this study, a cross-sectional based solid modelling method is proposed in order to guide the
user in decision making. The concept of ore-waste discrimination plays an important role in the
mineral resource estimation. Mine planning may be the subject of miscalculation when it is
made for mineral assets under threshold grade. The algorithm, which is capable of forming a
faster section in the case of different threshold grade values, provides a great convenience
especially for ore deposits which are difficult to model. However, this method should be used
only as a guide in decision making.
The generalized accuracy of the SVM algorithm presented in the study as a result of cross
validation can also be considered as an important criterion for the accuracy of the solid model.

REFERENCES

[1] NI43-101, National Instrument 43-101. 2011.

[2] Joint Committee of the Australasian Institute of, M., A.I.o.G. Metallurgy, and A. Minerals Council of,
Australasian code for reporting of mineral resources and ore reserves : (the JORC code), 1999 edition /
prepared by the Joint Ore Reserves Committee of the Australasian Institute of Mining and Metallurgy,
Australian Institute of Geoscientists and Minerals Council of Australia (JORC), ed. M. Australasian Institute
of, et al. 1999, [Gosford, N.S.W.?]: JORC.
[3] PERC, Pan-European Reserves & Resources Reporting Committee 2013.
[4] SAMREC, The South African Code for the Reporting of Exploration Results, Mineral Resources and Mineral
Reserves. 2016.
[5] Cortes, C. and V. Vapnik, Support-Vector Networks. Machine Learning, 1995. 20(3): p. 273-297.
[6] Platt, J.C., Fast training of support vector machines using sequential minimal optimization, in Advances in
kernel methods, S. Bernhard, et al., Editors. 1999, MIT Press. p. 185-208.
[7] Rifkin, R. and A. Klautau, In Defense of One-Vs-All Classification. J. Mach. Learn. Res., 2004. 5: p. 101-141.
[8] Bishop, C.M., Pattern Recognition and Machine Learning (Information Science and Statistics). 2006:
Springer-Verlag New York, Inc.

194