Computers & Geosciences 42 (2012) 71–78

Contents lists available at SciVerse ScienceDirect

Computers & Geosciences

journal homepage: www.elsevier.com/locate/cageo

Quantifying multi-element and volumetric uncertainty, Coleman McCreedy

deposit, Ontario, Canada
Ryan Goodfellow a,n, Francisco Albor Consuegra a, Roussos Dimitrakopoulos a, Tim Lloyd b
a
COSMO - Stochastic Mine Planning Laboratory, Department of Mining and Materials Engineering, McGill University, FDA Building, 3450 University Street,
Montreal, Quebec, Canada H3A 2A7
b
Vale, General Engineering Building, Hwy. 17 W, Door #10, Copper Cliff, Ontario, Canada P0M 1N0

a r t i c l e i n f o abstract

Article history: Traditional geostatistical modelling of orebodies and estimation of grade–tonnage curves do not
Received 30 March 2011 account for the uncertainty of the orebody grades and tonnages. Geological interpretations of complex
Received in revised form shapes are often over-constrained, and therefore do not properly identify the location of the ore. In
15 February 2012
these situations, tonnage is often under-estimated and grade is over-estimated, resulting in orebody
Accepted 21 February 2012
models used for mine planning that lead to costly financial decisions. This paper presents an approach
Available online 3 March 2012
aiming to better assess the uncertainty in an orebody model. The approach is applied to Vale’s West
Keywords: Orebody of the Coleman McCreedy Mine, a poly-metallic deposit containing nickel, copper, gold,
Wireframes platinum and palladium. To encapsulate the orebody’s variability and uncertainty, the nickel–copper
Multiple point statistics
sulphide mineralized zone is simulated using the single normal equation simulation method. The
Min/Max Autocorrelation Factors
realizations serve as the orebody models from which the grades of multiple elements are jointly
Volumetrics
Training image simulated using Min/Max Autocorrelation Factors. The final result is a series of equiprobable
representations of the mineralization that incorporates both grade and tonnage uncertainty. The case
study indicates that had conventional orebody estimations been used, there would have been a 10%
over-estimation of orebody volume, along with significant over-estimation of low-grade material and
under-estimation of high-grade material.
& 2012 Elsevier Ltd. All rights reserved.

1. Introduction and is based on extracting the so-called multiple point (MP)

statistics from a training image (Strebelle, 2002). Some recent
In cases where drillhole spacing is particularly tight and the developments on multiple-point methods include a list-based
geological domains are continuous, deterministic mapping of storage of patterns (Straubhaar et al., 2011), directly sampling
geological domain boundaries (i.e. a wireframe) is likely to be patterns from the training image (Mariethoz et al., 2010), filter and
sufficient in describing the mineralized volume of a deposit pattern-based simulations (Zhang et al., 2006; Arpat and Caers,
(Srivastava, 2005). To the contrary, deposits containing a mixture 2007) and high-order simulation (Mustapha and Dimitrakopoulos,
of two or more grade populations and uncertainty in the inter- 2010).
pretation of the edges of the geological domains lead to substantial Geological deposits typically contain several variables of interest
volumetric uncertainty, and stochastic treatment of the wireframes that are spatially correlated. The use of joint geostatistical techni-
offers a suitable alternative (Osterholt and Dimitrakopoulos, 2007). ques that maintain spatial correlation are not new (David, 1988;
Curvilinear geometries, typical in mineral deposits, are not prop- Goovaerts, 1997), however, the computational costs associated
erly modelled using traditional two-point spatial statistics such as with the simulation increase significantly with more variables,
variograms (Journel, 2007). The reproduction of geometries calls and require modelling of cross-correlations, which substantially
for the consideration of the joint categorical variability at three or increases with the number of variables being jointly simulated.
more points at a time. Strebelle (2002) proposed a stochastic A practical alternative to the ‘direct’ joint-simulation of variables is
simulation algorithm that does not require variogram modelling the decorrelation of variables introduced using principal compo-
nent analysis or PCA (David, 1988; Wackernagel, 2003). The
effectiveness of this approach is limited because PCA does not
n
Corresponding author. Tel.: þ1 514 398 4755x094748; fax: þ1 514 398 7099. eliminate cross-correlations at distances other than zero. To over-
E-mail addresses: [email protected] (R. Goodfellow),
[email protected] (F. Albor Consuegra),
come the limitations associated with PCA, minimum/maximum
[email protected] (R. Dimitrakopoulos), autocorrelation factors, MAF (Desbarats and Dimitrakopoulos,
[email protected] (T. Lloyd). 2000; Boucher and Dimitrakopoulos, 2009; Dimitrakopoulos and

0098-3004/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
doi:10.1016/j.cageo.2012.02.018
72 R. Goodfellow et al. / Computers & Geosciences 42 (2012) 71–78

Fonseca, 2003; Lopes et al., 2011) may be used to decorrelate a kriging system consisting of two-point covariances, the MP condi-
pertinent variables into spatially non-correlated factors that are tional distribution is conditioned to a single joint MP data event dn:
independently simulated and back transformed to correlated attri-
f ðx; sk 9dn Þ ¼ EfIðx; sk Þ9dn g ¼ PrfSðxÞ ¼ sk 9dn g, k ¼ 1, . . . ,K ð1Þ
butes. To further reduce the time required to simulate the deposit
for the variables of interest, MAF can readily be integrated with the where f ðx; sk 9dn Þ is the conditional probability that sðxÞ ¼ sk given
direct block simulation method (Boucher and Dimitrakopoulos, the data event dn and EfIðx; sk Þ9dn g is the conditional expectation of
2009). the indicator function Iðx; sk Þ given the data event dn, whereby
In the following sections, a general overview of multiple-point Iðx; sk Þ ¼ 1 if sðxÞ ¼ sk , otherwise Iðx; sk Þ ¼ 0.
simulation is given, followed by an overview of joint simulation of Let Ak denote the binary random variable indicating the
multiple correlated variables using Min/Max Autocorrelation Factors occurrence of data event dn at location x
at the block support scale. Both methods are then applied to a case 
1 if SðxÞ ¼ sk
study at Vale’s West Orebody at the Coleman McCreedy deposit, by Ak ¼ ð2Þ
0 otherwise
first simulating the ore envelope and subsequently simulating the
grades for gold, copper, nickel, palladium and platinum jointly Similarly, let D denote a binary random variable indicating
within the models. Following this, the results from the generated the occurrence of data event dn. The conditional probability of the
grade–volume curves are discussed in terms of volumetric uncer- node or location x corresponding to state sk is given by the
tainty and grade variability within the generated models. Finally, following simple indicator kriging expression:
conclusions from this case study are presented.
f ðx; sk 9dn Þ ¼ PrfAk ¼ 19D ¼ 1g ¼ EfAk g þ l½1EfDg ð3Þ

where EfDg ¼ PrfD ¼ 1g is the probability of the conditioning data

2. Multiple-point simulation of orebody models event dn occurring and EfAk g ¼ PrfSðxÞ ¼ sk g is the prior probability
for the state at location x is sk. Solving the simple kriging system
2.1. Definitions for the single weight l corresponding to data event dn leads to the
following solution (Guardiano and Srivastava, 1993):
Multiple-point statistics, or MP-statistics, consider a neighbour-
PrfAk ¼ 1,D ¼ 1g
hood of any n number of surrounding points jointly. The geometric f ðx; sk 9dn Þ ¼ PrfAk ¼ 19D ¼ 1g ¼ ð4Þ
configuration of a multiple-point data event dn centred at node A is PrfD ¼ 1g
called a template, tn , of size n. The size of the template, n, and its The numerator and denominator of Eq. (4) are inferred by
shape can be configured to capture any data events that inform the scanning the training image for both the number of replicates of
central node A. Fig. 1 shows an example of a data event on a the conditioning data event cðdn Þ, and the number of replicates
template with n¼4 (Osterholt and Dimitrakopoulos, 2007). ck ðdn Þ from the set of replicates cðdn Þ that have SðxÞ ¼ sk , i.e.
Specifically, consider an attribute S taking K possible catego-
ck ðdn Þ
rical states fsk ,k ¼ 1, . . . ,Kg, which may represent, for example, f ðx; sk 9dn Þ ¼ ð5Þ
cðdn Þ
lithological codes or metallurgical ore types. A geometric tem-
plate tn is defined by a set of n vectors fha , a ¼ 1, . . . ,ng. A data In the SNESIM (single normal equation simulation) algorithm
event dn of size n, centred at the location or node x, is defined by (Strebelle, 2002), the available conditioning data forming the data
the vectors in the template tn , offset from a central location x. event dn is stored in a search tree. The proportions for building the
The data event consists of a set of measured categorical state ccdf are retrieved by searching for similar data events in the search
values sðxþ ha Þ at locations ðxþ ha Þ, a ¼ 1, . . . ,n. Any permutation tree reading the related frequencies. In MP geostatistics, there is no
of data values sðxþ ha Þ, a ¼ 1, . . . ,n for the set of vectors fx þha , need to approximate the use of the global conditioning data event
a ¼ 1, . . . ,ng represents a distinct data event, hence will represent due to the use of a training image and the exact calculation of the
a different MP-statistic. probability distribution conditional to dn. It must be noted that given
Given that MP statistics are often used to characterize spatial that all MP statistics are derived from the patterns present in the
relationships of closely spaced data, training images (TIs) are used training image, the algorithm is said to be ‘‘training-image driven’’;
to supplement the multiple-point statistics that cannot be inferred the conditioning samples represented by, for example, drillhole data
from sparse drilling data alone. The geometries and underlying merely guide the algorithm to search for patterns from the training
patterns contained in a training image should be consistent with the image that fit the same data configuration dn. For this reason, it must
geological concept and interpretation of the deposit. be assumed that the training image is an accurate representation of
reality. For a more detailed discussion on the SNESIM algorithm, the
2.2. A conditional simulation algorithm for MP-geostatistics reader is referred to Remy et al. (2011). An overview of the general
steps of the simulation is given below:
Whereas traditional variogram-based conditional simulation
methods estimate a conditional distribution function by solving 1. Scan the training image and store occurrences of all data
events D.
2. Define a random path and visit all nodes to be simulated.
3. Simulate each node by:
a. Retrieving all data events containing the surrounding data
and previously simulated nodes.
b. Derive the local probability distribution from stored frequen-
cies of central values; the probability of finding a certain
categorical variable at the node given the surrounding data
event D is given by the Bayes’ relation for conditional
probability in Eq. (5).
c. Randomly select a pattern from the distribution and add
Fig. 1. Example of a data template. The central node A is to be simulated using the simulated node to the grid.
n surrounding known data events, D. 4. Repeat the previous steps for a desired number of simulations.
 R. Goodfellow et al. / Computers & Geosciences 42 (2012) 71–78 73

3. Joint simulation of correlated variables with Minimum/ points and the point data and the block data for the lth MAF
Maximum Autocorrelation Factors factor, respectively. The vector of N simulated values mlpn for each
factor l is obtained by solving the following system of equations
Consider a multivariate, stationary and ergodic spatial random similar to generating LU simulations with conditioning data
function of a continuous variable ZðxÞ ¼ ½Z 1 ðx, . . . ,Z p ðxÞT that has (Godoy, 2003; Boucher and Dimitrakopoulos, 2009)
been transformed to a standard Gaussian random function
mlpn ¼ LlpIV ðLlIIVV Þ1 mlIV þ Llpp wp ð12Þ
YðxÞ ¼ ½Y 1 ðx, . . . ,Y p ðxÞT ¼ ½fðZ 1 ðxÞÞ, . . . , fðZ p ðxÞÞT and is assumed
to be multi-Gaussian. Minimum/Maximum Autocorrelation Fac- where mlIV is the vector containing neighbouring point and block
tors (MAFs) are defined as the p orthogonal linear combina- data for factor l, wp is a vector of randomly generated standard
tions M i ðxÞ ¼ aTi YðxÞ,i ¼ 1, . . . ,p of the original multivariate vector normal numbers and LlpIV , LlIIVV and Llpp are obtained from the
YðxÞ. MAFs are derived assuming that YðxÞ is represented by a following Cholesky decomposition:
two-structure linear model of coregionalization (Desbarats and 2 l 3 2 32 lT T
3
Dimitrakopoulos, 2000). The MAF transformation can be rewrit- CIIVV ClpIV LlIIVV
l
0 6 LIIVV LpIV 7
4 5¼4 54 T 5 ð13Þ
ten as ClpIV Clpp LlpIV Llpp 0 Ll pp
MðxÞ ¼ ATMAF YðxÞ ¼ ATMAF /ðZðxÞÞ ð6Þ
The block value can be obtained by averaging the back-
and the MAF factors, AMAF , are derived from transformed points that discretize each block with Eq. (12).
AMAF ¼ Q 2 K1 The direct block MAF simulation algorithm (Boucher and
1 Q1 ð7Þ
Dimitrakopoulos, 2009) proceeds as follows:
where the eigenvectors Q 1 and eigenvalues K1 are obtained from
the spectral decomposition of the multivariate covariance matrix 1. Transform the data ZðxÞ to normal-score data YðxÞ.
B of YðxÞ at the zero-lag distance. More specifically, the spectral 2. Transform YðxÞ to MðxÞ using the MAF transformation in Eq. (6).
decomposition is computed as 3. Define a random path to visit each block to be simulated.
Q 1 BQ T1 ¼ K1 ð8Þ 4. For each block v, simulate the MAF factors at each of the
block’s N discretizing points mn ðxi Þ,i ¼ 1, . . . ,N with the LU
Q 2 is the matrix of eigenvectors from the following spectral decomposition using Eq. (12).
decomposition: a. Average the points mn ðxi Þ,i ¼ 1, . . . ,N over the block to
Q 2 ð12½Cz ðDÞT þ ½Cz ðDÞÞQ T2 ¼ K2 ð9Þ obtain mnV ðvÞ for further conditioning.
b. Back-transform the simulated nodes mn ðxi Þ,i ¼ 1, . . . ,N to
where the matrix Cz ðDÞ is an asymmetric variogram matrix at zn ðxi Þ and average the points xi A v to obtain the simulated
lag distance D for the regular PCA factors zðxÞ ¼ YðxÞA, where value at the block support scale, znV ðvÞ using Eq. (10). Write
1=2
A ¼ Q 1 K1 . K2 is the matrix of eigenvalues obtained from the the values to a file.
spectral decomposition in Eq. (9). In practice, the simplest way to 5. Repeat step (4) for each block v to be simulated.
ensure that the transformed variables are orthogonal is to test the 6. Repeat steps (3)–(5) to generate a set of multi-element
cross-variograms of the MAF factors for several lag distances, D, geological simulations.
and select the one which provides the best decorrelation; this
method permits applying the MAF transformation to the experi-
mental data set without having to derive theoretical variogram 4. Geology of the West Orebody
and cross-variogram models beforehand. For a more detailed
description of MAF for simulation of correlated attributes, the Vale’s West Orebody (WOB), located in Sudbury, Ontario,
reader is referred to Desbarats and Dimitrakopoulos (2000). Canada, is located 1463 m west of the Lower Coleman Orebody.
The direct block simulation method (Godoy, 2003), and more It has a known strike length of 213 m. The ore thickness ranges
specifically direct block simulation for multiple correlated vari- from 1.5 to 46 m with an average of 27 m. The upper portion of
ables (Boucher and Dimitrakopoulos, 2009), is used to simulate the orebody dips steeply to the south at 801 before flattening
the MAF factors MðxÞ independently for each of the blocks in the to 401 at depth. The orebody is composed of massive sulphides,
domain D A Rn . Simulation at the block-support scale using MAF stringer sulphides, and weaker disseminations within granite
calls for the block support vector random function ZV , which can breccia zones enclosed in the host footwall rocks. The host rocks
be expressed as are granite gneiss, granite breccia and Sudbury breccia.
1XN
ZV ðvÞ ¼ f1 ððATMAF Þ1 Mðxi ÞÞ, xi A v, 8i ð10Þ
Ni¼1 5. Simulation of the ore envelope
where each block v A D is discretized by N points at locations
xi A v and fi ¼ 1, . . . ,Ng denotes the indices for each of the 5.1. TI and hard data
discretization points. For each block to be simulated, the dis-
cretizing point values conditional to neighbouring point and A conventional deterministic orebody model is used to serve
block data are simulated with the joint LU simulation. Let ClIIVV as a training image for the simulation process. To generate the
be the covariance matrix for the conditioning data and ClIV be the model, the drillhole data is coded as inside (1) or outside (0) the
covariance matrix between the simulation grid’s points and the nickel–copper mineralized zone based on the presence low-grade
neighbouring point and block data for the lth MAF factor nickel. A wireframe envelope is generated and converted to a
2 3 block model on a regular grid, whereby all blocks that lie within
ClII ClIV h i the mineral envelope (wireframe) are designated as mineralized
l
CIIVV ¼ 4 l 5 and Cl ¼ ClpI ClpV ð11Þ
l pIV material (1) and all other blocks are designated as barren rock (0).
CVI CVV
It is noted that the low-grade nickel criterion, hence the wire-
where ClII , ClIV and ClVV are the point-to-point, point-to-block and frame envelope, is subject to interpretation; a single, smooth and
block-to-block covariance matrices for the lth MAF factor, respec- connected orebody model is generally preferred by the geologist.
tively; ClpI and ClpV define the covariances between the discretizing The conventional model is used as a training image and is
74 R. Goodfellow et al. / Computers & Geosciences 42 (2012) 71–78

comprised of 31  84  80 blocks with dimensions of 6.1  4.6 

4.6 m3. A cross-section of the training image is shown in Fig. 2(a).
The hard data used for the simulations is obtained from 15,319
drillhole samples composited over 1.52-m lengths.

5.2. Orebody simulation, specifics and post-processing

To assess the geological uncertainty in terms of volume and

tonnage, a set of 20 realizations of the orebody are generated
using the SNESIM algorithm (Strebelle, 2002; Osterholt and
Dimitrakopoulos, 2007). The simulations are generated on the
same grid as the conventional model (training image) defined in
Fig. 3. Indicator variogram reproduction of the post-processed simulations.
Section 5.1. Since this is a single orebody, connectivity for the
blocks is expected. In this application, the SNESIM algorithm
found it difficult to retrieve sufficient information related to the
connectivity and orientation of geological features in the orebody
in order to guarantee realizations that consisted of a single
completely connected orebody. One solution employed to over-
come this issue was to increase the number of nodes in the
template, thus providing more information to a node being
simulated. This, however, raises the computational costs for the
simulations.
Nevertheless, the connectivity issue was not completely over-
come, and the realizations required post-processing to remove
randomly allocated ore blocks in the model. A simulated anneal-
ing-based algorithm (Deutsch and Journel, 1998) is used to
post-process the simulations to better-reproduce the two-point
continuity of the training image. Osterholt and Dimitrakopoulos
(2007) also document similar issues related to connectivity and
smoothness when simulating iron ore channels. The ‘‘cleaned’’
realizations are referred simply as realizations in the subsequent
Fig. 4. Frequency distribution of the volume of ore for the post-processed SNESIM
sections. realization as a percentage of the volume of ore blocks in the training image.

5.3. Validation
and training image are significantly less than the variability in
Fig. 2 shows how the simulated realizations reproduce the the data. This behaviour is to be expected, given that the
large-scale features of the training image, such as the steeply algorithm is driven by the patterns available in the training image
dipping of the upper portion of the orebody and its flattening with rather than a variogram model; additionally, the post-processing
depth. The algorithm (Remy et al., 2011) allows the realizations to step enhances the training image’s variogram reproduction. The
reproduce the target waste proportion of 93.2%; the waste significant difference between indicator variograms suggests that
proportions of the 20 realizations fluctuate between 93.2% and there is a conflict between the data and the training image, and is a
94.1% with a mean of 93.9%. Fig. 3 shows the indicator variogram result of choosing a single, smooth and connected orebody model
reproduction for the SNESIM realizations, training image and rather than using a model that strictly obeys the low-grade nickel
data. The indicator variogram for vector h, gI ðh; sk Þ, is calculated coding applied to the drillholes. These issues have also been studied
using: by Osterholt and Dimitrakopoulos (2007) when simulating an iron
ore channel. To deal with such cases where there is a conflict
NðhÞ
1 X between the data and training image, data-driven methods, such as
gI ðh; sk Þ ¼ ½Iðxa ; sk ÞIðxa þh; sk Þ2 ð14Þ
2NðhÞ a ¼ 1 high-order simulation with spatial cumulants (Mustapha and
Dimitrakopoulos, 2010; Dimitrakopoulos et al., 2010), need to be
where NðhÞ denotes the number of pairs of points separated by the explored.
vector h, and the a is used to index these pairs. It is first noted
that the simulations accurately reproduce the indicator variogram of 5.4. Assessing volumetric uncertainty
the training image, however the variability for both the simulations
To assess the grade–tonnage uncertainty present in the
modelling of the deposit, the realizations generated by SNESIM
are used as distinct orebody models from which nickel, copper,
gold, palladium and platinum are jointly simulated. The fre-
45 m quency distribution, in terms of the percentage total volume of
ore in the simulation compared to the volume of ore in the
training image, is shown in Fig. 4. It is noted that the volume of
the ore blocks of the training image (a conventional deterministic
wireframe generated by a geologist) is approximately 10% higher
than the volume of the simulations. This highlights the need for
Fig. 2. Example of a data template. Cross-sections of (a) the training image,
the generation of multiple objective orebody models when
(b) cleaned maximum, (c) cleaned median and (d) cleaned minimum simulated assessing a mineral deposit. For the purpose of this study, the
models of the deposit (in terms of volume). simulated models with the maximum, median and minimum ore
 R. Goodfellow et al. / Computers & Geosciences 42 (2012) 71–78 75

volumes are retained for further joint multi-element simulation, score) variables; the variogram models of the MAF factors does
and are referred to as the Max, Med and Min orebody models. It is not necessarily correspond to the variography of the variables.
noted that in general, all simulations should be retained for joint
multi-element simulation. However, in this case there is very
6.3. Conditional simulation of MAF at the block support scale
little difference in terms of volumetrics for the majority of the
simulations (with the exception of the Max orebody model),
Conditional simulations for each of the factors in each of the
which is a result of the high drilling density within the miner-
orebody models are performed independently using Eq. (12) and
alized zone. Fig. 2 shows sample cross-sections of the selected
averaged into blocks using Eq. (10); the point support informa-
orebody models.
tion is retained for a variogram validation with the drillholes.
The simulations are generated inside the orebody models using
6.1  4.6  4.6 m3 blocks and a 4  3  3 node discretization
6. Joint simulation of nickel, copper, gold, palladium and
density on the same grid as the wireframe, resulting in
platinum
502,524, 446,940 and 436,536 nodes within the limits of the
maximum, median and minimum simulated orebody models,
6.1. Normal score and MAF transformation
respectively. Ten simulations for each of the factors are generated
for each of the three simulated orebody models and are validated
Similar to traditional orebody estimation methods, the drill-
in detail for reproduction of data, histograms and variograms
holes are filtered for each orebody model, such that only the
on the point support. Fig. 5 shows a cross-variogram between
drillholes inside the model are kept. Prior to applying the MAF
the first two MAF factors for the median model. As one would
transformations to the spatially correlated variables, each of
expect during simulation, there are no spatial cross-correlations
the elements is transformed to a normal distribution. Normal
between the two factors, hence the MAF factors are indeed
score transformations are based on rank ordering of the data and
independent and have been successfully decorrelated. The
decrease the influence of outliers. The transformation matrix AMAF
remaining cross-variograms are very similar to the one presented,
(Eq. (7)) for the median orebody is shown in Table 1. MAF are
thus are omitted. A validation of variograms, cross-variograms
calculated by multiplying the vector of Cu, Ni, Au, Pd and Pt by a
and histograms is presented only for the data space in the
vector of loadings from the rows of the transformation matrix.
subsequent section.
The D lag in Eq. (9) used in this example is 9 m, and is derived
experimentally by testing several lag distances to assure a
suitable decorrelation and stable MAF decomposition. 6.4. Validation of results

Validation of the jointly simulated variables involves calcula-

6.2. Variography of MAF
tion of histograms, experimental variograms and cross-vario-
grams of the simulated point realizations in the data space to
The variogram models used for simulating each of the MAF
ensure reproduction of original data and their spatial character-
factors in the median orebody model are presented in Table 2.
istics. Fig. 6 shows the variograms and cross-variograms for the
Each of the variogram models was assumed to contain a nugget
original drillhole data and conditional simulations for the median
and a spherical structure. It is noted that MAF variograms are
orebody at the point support scale. All results suggest that the
linear combinations of the variograms of the original (normal
reproduction of the original data spatial characteristics by the
simulated realizations is reasonable. Recall that the variograms
Table 1 and cross-variograms of original variables are not directly used in
MAF factor loadings described as linear combinations of the input variables for the the joint simulation based on MAF, which used the variograms
median (Med) orebody model. of the independent MAF. Fig. 7 shows the reproduction of the
histograms of the 10 simulations and the original drillhole data
Metal MAF 1 MAF 2 MAF 3 MAF 4 MAF 5
for Au, Cu, Ni, Pd and Pt in the median orebody. All graphs
Au 0.063  0.292 0.181  0.009  0.054 indicate that the simulations reproduce well the initial drillhole
Cu 0.783 0.210 0.012 0.240 0.154 data. Fig. 8 shows sample cross-sections for each of the simulated
Ni 0.296 0.015 0.038 0.168 0.391
metals inside the Min, Med and Max orebody models. It is noted
Pd 0.539 0.917 0.940 0.731 0.762
Pt  0.070 0.173 0.287 0.617 0.490
that by inspection, the grades are similar in all of the models,
however there are some fairly significant differences in the
shapes of the models, which highlights the benefits of generating
Table 2 a series of equiprobable orebody models to obtain a better
Variogram models used to simulate the MAF factors for the median (Med) understanding of volumetric variations.
orebody.

Variable Variogram model Major (x) Minor (z)

AZ1 DIP1 AZ1 DIP1

 
MAF 1 h hy hz 90 0 0 90
gðhÞ ¼ 0:54 þ 0:46sph x , ,
19 15 6
 
MAF 2 h hy hz 135 0 45 168
gðhÞ ¼ 0:48 þ 0:52sph x , ,
43 30 17
 
MAF 3 h hy hz 135 0 45 68
gðhÞ ¼ 0:56 þ 0:44sph x , ,
37 26 10
 
MAF 4 h hy hz 135 0 45  45
gðhÞ ¼ 0:41 þ 0:59sph x , ,
44 44 44
 
MAF 5 h hy hz 225 45 225  45
gðhÞ ¼ 0:41 þ 0:59sph x , , Fig. 5. Example of a cross-variogram between MAF 1 and MAF 2 for all simula-
51 40 17
tions from the median orebody.
76 R. Goodfellow et al. / Computers & Geosciences 42 (2012) 71–78

Fig. 7. Histogram reproduction of the 10 simulations and original drillholes for Ni,
Cu, Au, Pd and Pt in the median orebody model.

7. Risk analysis for decision making

To better understand the relationship between volumes and

grades for the deposit, Fig. 9 shows the changes in volume for a
nickel cutoff grade for the 6.1  4.6  4.6 m3 blocks. It is noted that in
the graph, the volumes above a cutoff for the Min and Med orebodies
are very close, and it is difficult to distinguish between the sets of
simulations. The volumes above a certain cutoff grade are signifi-
cantly higher for the Max orebody simulations, which is to be
expected given that the Max model has a 10% larger volume than
the Med model (whereas the Min model is only 2% smaller than the
Med model). This trend is particularly apparent for low cutoff grades
for all of the metals; at higher cutoffs, all of the curves for the Min,
Med and Max orebodies converge.
As a basis for comparison, the nickel grade is estimated within the
training image using ordinary kriging. The volume–cutoff curve
for this model (Fig. 9) highlights the importance of incorporating
both volumetric and grade uncertainty into the geology models. It is
clear that in this case study (however not necessarily in general),
neglecting to incorporate the volumetric uncertainty would lead to a
significant over-estimation of the orebody volume, as can be seen at
low cutoffs in Fig. 9. This figure also highlights the importance of
stochastic simulation methods over conventional grade estimation
methods, which would have led to a model that overestimates the
low-grade material and underestimates the high-grade material. It is
clear that using the traditional geostatistical methods, comprised of
creating a subjective and deterministic orebody model and estimating
within that model may lead to costly business decisions in the mine
planning phase.

8. Conclusions

Fig. 6. East–West variogram and cross-variogram reproduction for the median orebody This paper presents an approach to assess volumetric and multi-
(note that the heavy line is the drillhole variogram and the light lines are simulations). ple-element uncertainty, and is applied at West Orebody of Vale’s
 R. Goodfellow et al. / Computers & Geosciences 42 (2012) 71–78 77

Fig. 8. Volume above Ni cutoff grade curves for the simulated orebodies with 6.1  4.6  4.6 m3 blocks.

Coleman McCreedy deposit, located in Sudbury, Ontario, Canada. In geostatistics method, was first used to simulate a series of equiprob-
order to better quantify the spatial relationships, curvilinear features able orebody models. A conventional, deterministic orebody model is
and volumetric uncertainty in the orebody, SNESIM, a multiple-point used as a training image to provide a set of patterns for the simulation
78 R. Goodfellow et al. / Computers & Geosciences 42 (2012) 71–78

David, M., 1988. Handbook of Applied Advanced Geostatistical Ore Reserve

Estimation. Elsevier, Amsterdam 232 pp.
Desbarats, A.J., Dimitrakopoulos, R., 2000. Geostatistical simulation of regionalized
pore-size distributions using Min/Max autocorrelation factors. Mathematical
Geology 32 (8), 919–942. doi:10.1023/A:1007570402430.
Deutsch, C.V., Journel, A.G., 1998. GSLIB: Geostatistical Software Library: and
User’s Guide, 2nd ed. Oxford University Press, New York, NY 369 pp.
Dimitrakopoulos, R., Fonseca, M., 2003. Assessing risk in grade-tonnage curves in a
complex copper deposit, northern Brazil, based on an efficient joint simulation
of multiple correlated grades. In: Proceedings of the Application of Computers
and Operations Research in the Minerals Industries, Cape Town, South Africa,
pp. 373–382.
Dimitrakopoulos, R., Mustapha, H., Gloaguen, E., 2010. High-order statistics of
spatial random fields: exploring spatial cumulants for modeling complex
non-Gaussian and non-linear phenomena. Mathematical Geosciences 42 (1),
65–99. doi:10.1007/s11004-009-9258-9.
Godoy, M., 2003. The Effective Management of Geological Risk in Long-Term
Scheduling of Open Pit Mines. Ph.D. Dissertation. University of Queensland,
Brisbane, Australia, 256 pp.
Goovaerts, P., 1997. Geostatistics for Natural Resources Evaluation. Oxford Uni-
versity Press, New York, 496 pp.
Guardiano, F., Srivastava, M., 1993. Multivariate geostatistics: beyond bivariate
Fig. 9. Volume above Ni cutoff grade curves for the simulated orebodies with moments. In: Geostatistics-Troia. Kluwer Academic, Amsterdam, pp. 133–144.
6.1  4.6  4.6 m3 blocks. Journel, A.G., 2007. Roadblocks to the evaluation of ore reserves—the simulation
overpass and putting more geology into numerical models of deposits. 2nd ed.
In: Dimitrakopoulos, R. (Ed.), Orebody Modelling and Strategic Mine Planning,
Spectrum Series, vol. 14. The Australasian Institute of Mining and Metallurgy,
algorithm. Given that there is a conflict between the data and the
pp. 29–32.
training image in terms of spatial correlations and proportions, the Lopes, J.A., Rosas, C.F., Fernandes, J.B., Vanzela, G.A., 2011. Risk quantification in
majority of the simulated orebody models require post-processing grade–tonnage curves and resource categorization in a lateritic nickel deposit
to enforce the geologists’ connectivity requirement in the orebody. using geologically constrained joint conditional simulation. Journal of Mining
Science 47 (2), 166–176. doi:10.1134/S1062739147020043.
A subset of the simulations, specifically the orebody models that
Mariethoz, G., Renard, P., Straubhaar, J., 2010. The direct sampling method to
represent the highest, intermediate and minimum quantities of perform multiple-point geostatistical simulations. Water Resources Research
mineralized material, are then retained for joint simulation of nickel, 46 (W11536), 1–14. doi:10.1029/2008WR007621.
copper, gold, platinum and palladium using Min/Max Autocorrelation Mustapha, H., Dimitrakopoulos, R., 2010. High-order stochastic simulation of
spatially distributed natural phenomena. Mathematical Geosciences 42 (5).
Factors and Direct Block Simulation. This method permits the efficient
doi:10.1007/s11004-010-9291-8.
joint simulation of the spatially correlated metals without having to Osterholt, V., Dimitrakopoulos, R., 2007. Simulation of wireframes and geometric
fit cross-variogram models for each pair of elements. The result is a features with multiple-point techniques: application at Yandi iron ore deposit.
set of equiprobable orebody models that account for both volumetric 2nd ed. In: Dimitrakopoulos, R. (Ed.), Orebody Modelling and Strategic Mine
Planning, Spectrum Series, vol. 14. The Australasian Institute of Mining and
and multi-metal uncertainty. Results indicate that had a conventional,
Metallurgy, pp. 51–59.
deterministic orebody model been used with conventional grade Remy, N., Boucher, A., Wu, J., 2011. Applied Geostatistics with SGeMS: A User’s
estimation methods, there could have been as much as 10% over- Guide. Cambridge University Press, New York, 284 pp.
estimation of low-grade mineralized material. In order to alleviate the Srivastava, R.M., 2005. Probabilistic modeling of ore lens geometry: an alternative to
deterministic wireframes. Mathematical Geology 37 (5), 513–544. doi:10.1007/
issues with conflicts between training images and data, future work s11004-005-6670-7.
will focus on data-driven high-order stochastic simulation methods Straubhaar, J., Renard, P., Grégoire, M., Froidevaux, R., Besson, O., 2011. An
for simulating indicator variables (orebody models) and joint simula- improved parallel multiple-point algorithm using a list approach. Mathema-
tion for multiple elements. tical Geosciences 43 (3), 305–328. doi:10.1007/s11004-011-9328-7.
Strebelle, S., 2002. Conditional simulation of complex geological structures using
multiple-point statistics. Mathematical Geology 34 (1), 1–21. doi:10.1023/
References A:1014009426274.
Wackernagel, H.J., 2003. Multivariate Geostatistics. Springer-Verlag, Berlin,
256 pp.
Arpat, G.B., Caers, J., 2007. Conditional simulation with patterns. Mathematical
Zhang, T., Switzer, P., Journel, A.G., 2006. Filter-based classification of training
Geosciences 39 (2), 177–203. doi:10.1007/s11004-006-9075-3.
image patterns for spatial simulation. Mathematical Geosciences 38 (1),
Boucher, A., Dimitrakopoulos, R., 2009. Block simulation of multiple correlated
63–80. doi:10.1007/s11004-005-9004-x.
variables. Mathematical Geosciences 41 (2), 215–237. doi:10.1007/s11004-
008-9178-0.