Mining Geostats Mps
Mining Geostats Mps
Mining Geostats Mps
SUMMARY
In open pit mining, the information collected during production is abundant and generally of good quality. Using
a massive historical production information from a hydrothermal gold deposit, a realistic training image (TI) of
the actual mineralisation has been developed in the mined-out area of the open pit. The objective of developing
the training image in the context of Multiple Point Simulations (MPS) is to provide an analogue to simulate
deposits in similar geological settings as well as exploring deeper extensions of the same deposit.
In this application, the training image is used also as a reference for comparison with the simulated images. A
case study is presented and it is based on a complex geological system of a hydrothermal deposit characterised
by a number of distinct facies (pod zones) with high local variability with non-stationarity.
Firstly, 50 mineralisation facies simulations have been performed, this is followed by a second stage 100 block
gold grade simulations. The grade simulations are repeated for each facies independently with adapted
parameters characterizing the statistical distribution and the variograms. The facies and grade simulations are
finally merged by means of a cookie cutting procedure to get 100 realizations of the Au grade.
Two sets of boreholes corresponding to different drilling patterns of inclined boreholes typical of feasibility or
new mining mineral resource definition drilling have been used to condition the simulations. The statistical
analysis on the distributions of facies proportions and local statistics on similarities with the reality (i.e. the TI)
are provided. The capability to delineate the mineralized areas is also addressed. Comparing E-type estimates
calculated from the simulated gold grades with the actual in situ production grades based on grade control
drilling allows quantifying the uncertainty associated with the different sampling patterns used for the
alternative techniques.
J.DERAISME AND W. ASSIBEY-BONSU
INTRODUCTION
Simulations of ore bodies have multiple applications for the mining industry. The most important ones are linked
to uncertainty assessment issues and sampling optimization. Geology and grade have to be simulated
consistently with specific methods. Due to the high variability associated with the input parameters, eg, grade,
the choice of the method for simulating the geology is critical as simulated models will reproduce quite different
features depending on the method (eg, Object based simulations, Sequential indicator simulations, Truncated
Gaussian simulations). In the case study presented in this paper, a geological model with sufficient level of detail
and a high degree of confidence can be obtained from grade control data at the production stage, which provides
a reliable training image to be used in the Multiple Point Simulation process (MPS) (Strebelle 2002).
In addition to the training image, conditioning data from sampling of existing boreholes were used to control the
simulations of the geology. Grade simulations using turning bands method and conditioned by the same
boreholes were then carried out to obtain realizations of grades for each geological facies. The comparison of the
differences between simulated grades and grades of the training image, considered as the ground truth, allows
quantifying the uncertainty as well as the precision and efficiencies based on the different sampling campaigns.
Based on the orientation study done within the mined out area of the pit, and based on recommended drilling
pattern in the deeper part of the deposit, an E-Type block estimates as well as facies models would be
developed from MPS/simulations in the same manner, particularly using the same training image from the
mined out area of the deposit.
METHODOLOGY
Some Theoretical Considerations for Block Simulations Used in the Case Study
The Discrete Gaussian Model (DGM) (Chils and Delfiner 2012) was used to generate directly block values,
without going through an average of n points discretising the block.
Under the hypotheses of the Discrete Gaussian Model (DGM) we can simulate directly block values (Deraisme
et al. 2008). This method is very efficient as it saves a lot of computing time compared to the traditional
approach where the block values are obtained by averaging simulated points by discretizing the blocks. The key
hypothesis in the methodology considers that a point randomly located within a block is on the average equal to
the block value (Cartiers relationship). This property allows us to link point and block values by means of a
linear regression between their Gaussian transforms Yx and Yv. It is illustrated by the schematic figure below
(Figure 1).
Figure 1: Scheme of the change of support by means of Gaussian anamorphosis between raw variable Z
and its normal score transform Y.
The slope of the linear regression is the coefficient of correlation between both Gaussian variables. It is the so-
called change of support coefficient r, it has been demonstrated (Emery and Ortiz 2005) that the square of r is the
dispersion variance of the Gaussian variable regularized on the block (Y(v)):
2
A MINING APPLICATION OF MULTIPLE POINT SIMULATIONS
The block Gaussian variable Yv is nothing but the Gaussian variable regularized on the block Y(v) normalized by
the change of support coefficient:
For the direct block simulation method the input variogram model is obtained from the regularized variogram of
the point Gaussian variable by scaling it by the square of the r coefficient:
100 Grade simulations are generated using the turning bands simulation algorithm. A change of support model
of 5mx5mx3m blocks was obtained from a Gaussian anamorphosis model applied to composited data and the
variogram model of gold grade of each facies population.
Both facies and grade models in the case study were merged. For each block and each simulation, the
simulated grade assigned to that block and the simulation of same rank is the simulated grade of the simulated
facies of same rank. For example if the simulation #1 gives the facies 1 and the simulation #2 gives facies 3,
the simulation # 1 will have a grade from the grade simulation #1 of facies 1 and the simulation # 2 will have a
grade from the grade simulation #2 of facies 3.To obtain get 100 grade simulations from only 50 facies
simulations the following rule has been applied, 2 grades have been assigned to each facies simulation, i.e.
grade rank equal to the facies rank and equal to the facies rank + 50.
CASE STUDY
Geology
The mine exploits oxide and fresh hydrothermal mineralization located in Tarkwaian sediments. The
mineralization occurs in a Banket Sandstone formation split into 2 main Fault blocks. In each block gold
mineralisation is concentrated in 3 to 4 facies also called pod zones with different characteristics in terms of
grade distribution and variograms.
The model of gold concentration is illustrated in Figure 2.
3
J.DERAISME AND W. ASSIBEY-BONSU
A numeric model of the mined out deposit is available at a fine resolution 5mx5mx3m with rocktype and
mineralisation codes (or facies). This model is used for two purposes:
as a training image driving the MPS simulations,
as a ground truth reference or reality, i.e. actuals to which the different estimations and techniques
were compared for their validity and relative effciencies.
4
A MINING APPLICATION OF MULTIPLE POINT SIMULATIONS
Figure 3: Perspective view of the training image of the Fault block 2 with 3 pod zones.
Table 1shows the comparisons of the pod zones proportions between the training image and the sampling based
on the 2 boreholes data sets. Significant statistical differences are observed on the pod zones proportions
between the training image and that estimated from the sampling of the 2 boreholes data sets and especially for
the 40x80m grid.
Table 1: Statistics of the number of data in each pod zone of the Fault Block 2 from the training image
and from the sampling by boreholes on 40mx80m and 20mx40m patterns. The complement to 100% is
waste.
Fault Block 2
(11.6%)
(6.1%)
(1.85%)
MPS Results
Figure 4 shows an example of the simulation of one fault block conditioned by the boreholes of the pattern
20mx40m. It demonstrates a good reproduction of the general shape of the facies with higher variability at
small scale.
5
J.DERAISME AND W. ASSIBEY-BONSU
Figure 4: Perspective view of one simulation of the pod zones of the Fault block 2.
The distribution of the pod zone volumes from simulations conditioned by the respective sets of bore holes
(Figure 5), is consistent with the expectation of a significant reduction of the variance around the mean value
when more boreholes are used as input. The variance may be interpreted as the variance of the estimation error
on the volume since the average of simulations can provide a non-biased estimate and each simulation is an
equi-probable representation of the unknown.
Thus, the information provided by the simulations is the quantification of the error magnitude for the different
sampling patterns.
In addition to the error variance, the confidence interval at any risk level can be calculated experimentally
which is critical for capital intensive mining projects.
Table 2 and Figure 6 provide the tonnages calculated from the training image which are the actual tonnages,
since the training image is considered to be the reality. It can be observed that the risk of bias is important, when
a pod zone represents a small proportion of the total volume. This just points out that MPS does not guarantee
that the simulated facies proportions will match those of the training image, hence the difficulty of estimating
volumes without bias from sparse drilling. However, the denser grid ie 20mx40m) provides a more practical
estimate. The average MPS simulated tons for all the facies combined based on the 20mx40m grid is within +/-
8% standard deviation with a probability of 90% (See Table 2).
Frequencies
Frequencies
0.075 0.075
0.10
0.050 0.050
0.05
0.025 0.025
0.20 0.15
Frequencies
Frequencies
Frequencies
0.10
0.15
0.10
0.10
0.05
0.05
0.05
Figure 5: Histograms of the number of simulated blocks for each pod zone of the Fault Block 2 (from the
50 simulations with sampling pattern 20mx40m at the top and from the 50 simulations with sampling
pattern 40mx80m at the bottom).
6
A MINING APPLICATION OF MULTIPLE POINT SIMULATIONS
Table 2: Actual Tonnage, mean simulated Tonnage and Confidence Intervals at the risk level of 90% from
50 simulations of the different pod zones.
Actual Tonnage q5-q95 Tonnage
Mean Tonnage (Mt)
(MT) Sampling (Mt)
20x40 7,24 1,5
RW2 6,99
40x80 7,5 1,87
20x40 2,82 0,49
RE2 3,73
40x80 2,89 1,24
20x40 0,9 0,48
RS2 1,14
40x80 0,63 0,64
20x40 5,96 0,58
RW3 7,33
40x80 6,89 2,86
20x40 1,62 0,24
RE3 1,66
40x80 1,2 0,6
20x40 0,75 0,18
RS3 0,74
40x80 0,36 0,73
20x40 3,88 0,35
RF3 3,48
40x80 3,47 1,48
20x40 23,17 3,94
TOTAL 25.07
40x80 22,94 9,7
Figure 6: Comparison of average tonnages per facies with q5-q95 intervals from simulations conditioned
by both drillholes patterns.
Table 3: Probabilities of blocks to be in same pod zone as in the training image (left columns) and to be in
another pod zone (right column).
Block in SAME facies as in TI Block in OTHER facies than in TI
7
J.DERAISME AND W. ASSIBEY-BONSU
1.25
Direction East-West Direction North South Direction Vertical
1.25
1.5
1.00
1.00
0.75
1.0
0.75
0.50
0.50
0.5
0.25
0.25
Figure 7: Example of variograms on Gaussian grade for one pod zone: the double line is the model input
in the simulation process, the single line is the average of the simulated variograms , the dotted lines show
the envelope of the simulated variograms.
After having populated the simulated models with gold (Au) grades, some simple statistics can be calculated
like E-type estimates, standard deviation between simulated values and actual values. The bias observed in
the facies simulations when the data are too sparse (boreholes patterns of 40mx80m) is still very pronounced on
grades (Table 4).
Table 4: Average tonnages and grade before and after cut-off from 100 grade simulations and actual
figures.
20x40 21.37
The actual
21.93-25.87 2.22
grade profile
0 cutoff per level can
40x80 21.56 18.89-28.59 1.84
be plotted
Actual figures 22.26 2.24 and
compared
20x40 19.78 18.21-21.5 2.55 with the E-
after cutoff type
40x80 20.72 16.84-25.04 1.93 estimate of
gold grade
Actual figures 21.43 2.31
calculated
from the 100 simulations and the percentiles at 5% and 95% risk. Figure 8 shows an acceptable match.
8
A MINING APPLICATION OF MULTIPLE POINT SIMULATIONS
2,7
2,5
2,3
2,1
1,9
1,7
1,5
1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47
Figure 8: Vertical profile of actual grade, simulated mean grade and percentiles at 5% and 95% for each
level 3m (level 1 at the pit bottom).
CONCLUSIONS
The application of MPS to mining data has proved to be efficient to produce images of geological variables
characterized by their extreme non stationarity. The main historical challenge of MPS application which is often
pointed out, relates to deriving geological training images, which does not apply to the case study presented in
this paper, as the training image in this case has been obtained from a massive production grade control data. In
this particular application the training image was used twice:
to calculate probabilities of different spatial patterns between different facies
to provide a reference for comparison with the simulated outcomes.
The MPS methodology aims at reproducing geological shapes and transitions between facies at different scales
but there is no guarantee that the statistics of the facies proportions will match those of the training image
especially when only limited drilling data is available for the MPS modeling which could be the case for very
early feasibility mining projects. The study showed that the comparison with the training image makes sense
especially on a local scale, only if a sufficient number of drillhole data is available for conditioning the
simulations.
The study further showed that when a reliable geological/mineralisation training image is available and adequate
representative sample data have been drilled in similar geological unmined extension areas, MPS and block
simulations could assist in providing facies tonnages and grade models for feasibility/ new areas where mining
is yet to be extended. MPS could also assist with sample pattern optimization when a training image is readily
available. . Though computer time has remained historically a bottleneck, currently, this seems to being
overcome as the required results are achievable timeously with most of software implementation of the
simulation algorithms. Additional mining application case studies of MPS is further recommended as the use of
the methodology has important applications for new exploration projects as well as mining extensions into
similar geological orebody settings.
ACKNOWLEDGEMENTS
The authors are grateful to Gold Fields Ltd for the permission to publish this paper based on a real gold mine.
REFERENCES
Chiles, J.P. and Delfiner P., 2012, Geostatistics modeling spatial Uncertainty, Wiley series in probability and
statistics, pp 449-464.
Deraisme J, Rivoirard J, Carrasco P, 2008. Multivariate Uniform Conditioning and Block Simulations with
Discrete Gaussian Model : Application to Chuquicamata deposit, in Proceedings 8th International Geostatistics
Conference, 2008, Santiago, Chile:69-78.
9
J.DERAISME AND W. ASSIBEY-BONSU
Emery, X. and Ortiz J. 2005, Internal Consistency and Inference of Change-of-Support Isofactorial Models,
Geostatistics Banff 2004, Springer, 2005, p. 1057-1066.
Strebelle S (2002) Conditional simulation of complex geological structures using multiple-point statistics. Math
Geol, 34: 121
10