Professional Documents
Culture Documents
Preprint 99-18
OPTIMAL GRADE CONTROL USING GEOSTATISTICS AND ECONOMICS:
METHODOLOGY AND EXAMPLES
C. V. Deutsch
Univ. of Alberta
Edmonton, AB, Canada
E. Magri V.
Univ. of Chile
Santiago, Chile
K. Norrena
Univ. of Alberta
Edmonton, AB, Canada
mean 1.340
std. dev. 0.342
1; if z > zc coef. of var 0.255
( ; c) =
i z z (5) maximum 3.806
0; otherwise upper quartile
median
1.530
1.310
lower quartile 1.090
Frequency
called ore and ,cw if the material is called waste. Never-
theless, the notion of lost opportunity is very real; it does
cost money to mistakenly put high grade ore on the waste 0.04
dump.
This numerical approach is very flexible. It is straight-
forward to consider complicating factors in the calculation 0.0
of Pore and Pwaste . 0.0 1.0 2.0 3.0
grade
Numerical Laboratory
The concept of calculating Pore and Pwaste and then choos-
Figure 1: Histogram of exhaustive distribution of true grades.
ing the maximum profit selection (MPS) is simple and con-
sistent with engineering and economic principles; however,
we have to check that the results are significantly better of truth models have been considered with spatial hetero-
than conventional practice (kriging). Checking with real geneity that mimics real deposits. Kriging and simulation
data is notoriously difficult since the underlying true grades methods are given the same information to make fair com-
are never known and the grade of the material placed on parisons. Finally, we have attempted to enumerate all of
the waste dump is poorly known. We will use Monte Carlo the practical considerations (such as free selection, practi-
simulation as a Numerical Laboratory for testing different cal dig limits, blast movement, etc) that would make theo-
methods. retical revenue gains impracticable.
The experimental steps in our numerical laboratory con-
sist of:
AN EXAMPLE
1. Build a fine scale true grade model (together will all For illustration, a 2-D example from a gold mine will be
coregionalized variables), used. Although the grades and all economic parameters
2. Sample blasthole grades from the truth model at some are fictitous, they have been chosen to mimic the features
realistic spacing, of a real mining operation. It would have been easy to show
an example with real data; however, there would have been
3. Add sampling (and perhaps location) error to the “true” no underlying true distribution of grades to compare the re-
blasthole grades, sults. There is nothing special about using gold in this ex-
ample. In fact, more practical applications of this method-
4. Apply methods M1 ; M2 ; : : : ; Mm to establish differ- ology (to date) have been to base metal mines.
ent ore / waste classification, The economic parameters: milling cost ct =$12 :00=t,
5. Calculate the revenue R1 ; R2 ; : : : ; Rm generated by
ore mining cost c$1 o = :00=t , waste mining cost cw =
$1:00=t, recovery 0r =:80 , price $12 p = :00=g . The
each method, and then, breakeven cutoff grade may be calculated as 1:25g= t. The
6. Repeat for different truth realizations, models of spa- distribution of gold grades is illustrated on Figure 1.
tial correlation, and sampling practices to establish An important advantage of a synthetic example is the
the domain of applicability of each method, that is, ability to look at different types of deposits with different
where Rj Ri ; 8i 6= j . levels of sampling error. The first spatial distribution of
grades we will consider is shown on Figure 2. Note the N-
We have considered a 2-D truth model representing a single S anisotropy and continuous regions of high and low grade.
bench. There is no limitation to consider a single bench; in This 2-D example is 200 m E-W by 300 m N-S.
fact, all methods should perform somewhat better having Blastholes were taken on a 10 m grid. Figure 3 shows
data from the bench above and exploration drillhole data. a location map of the blastholes and Figure 4 shows a his-
Although well established in engineering studies, there togram. In practice, each individual blast will be much
are a number of concerns with such a Monte Carlo proce- smaller and all of the blasthole grades will not be available
dure: (1) the truth models are often too simplistic, (2) any at the same time. Also, in practice, there will be blasthole
method Mj that makes use of the underlying random func- grades available from the bench above the current working
tion model used to generate the truth model could appear bench. These issues have no affect on the “fairness” of our
unrealistically good (Rj too high), and (3) there may be comparison or the illustrative nature of the example.
practical considerations that make it impossible to achieve The blasthole grades are not perfect. The initial base
any modelled revenue gains. case considered a 10% relative error uniformly distributed
Awareness of these concerns help design the experi- with a zero average. Figure 5 shows a cross plot of the true
mental procedure used in the numerical laboratory. A range grades and the blasthole samples. This error level is realis-
4 Copyright (c) 1999 by SME
Blasthole Grades
Number of Data 600
0.12 mean 1.338
std. dev. 0.351
coef. of var 0.262
maximum 2.564
upper quartile 1.523
median 1.285
lower quartile 1.089
True Grades 0.08
minimum 0.410
Frequency
300
3.0
0.04
2.5
2.0
0.0
0.0 1.0 2.0 3.0
North
0.0
0
0 East 200
True Blasthole Grade 2.0
Figure 2: Gray-scale map of true grades for example application.
1.0
1.0
0.8
3.0
γ 0.6
0.4
2.0
0
0 East 200
North
0
0 East 200
to $864k.
Tweleve different truth models are considered to estab-
lish the generality of these results, see Table 1. Each truth
model leads to slightly different maximum revenue (recall
that the maximum was $1,011k in the base case). To make
the results comparable the revenues from MPS and kriging
have been normalized by the maximum revenue divided
by $1,000k; thus, the maximum revenue is $1,000k in all
cases. The same blasthole spacing and sampling error has
0
been considered in all cases; only the nugget effect and 0 East 200
range of correlation have been changed. MPS systemat-
ically leads to greater revenue than kriging, which is not Figure 12: Map of ore / waste classification based on maximum
surprising since the objective of MPS is to maximize profit profit selection.
whereas the goal of kriging is to create an estimate with
minimum squared error (see paper Minimum Variance or
Maximum Profitability by Srivastava, 1987).
Figure 13 shows a plot of the revenue of MPS and krig-
7 Copyright (c) 1999 by SME
MPS Kriging Relative Ranges (X/Y)
Based Nugget
$ 924,000 $ 855,000 0.0 50.0 / 50.0
Revenue $k
$ 789,000 $ 756,000 0.0 15.0 / 15.0
$ 853,000 $ 808,000 0.0 20.0 / 20.0
$ 946,000 $ 872,000 0.0 100.0 / 100.0
kriging-based
$ 955,000 $ 880,000 0.0 150.0 / 150.0 400.
$ 960,000 $ 884,000 0.0 200.0 / 200.0
200.
Table 1: Tabulated revenue achieved by grade control for differ-
ent underlying random function models. The maximum revenue
achievable with perfect selection is $1,000,000 in each case.
0.
0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8
ing versus the relative nugget effect of the underlying grade Relative Nugget Effect
models. The revenue attained by both methods decreases Figure 13: Chart of revenue (1000’s $) versus relative nugget
as the nugget effect increases. Beyond a nugget effect of effect with the simulation-based approach to grade control and
0.7 the predictability of the grades is so poor that the rev- the kriging-based approach to grade control.
enue is negative (mining the entire bench a either ore or
waste would lead to a loss). The MPS procedure leads to
significantly greater revenue when the nugget effect is less
than 0.7.
Figure 14 shows a plot of the revenue of MPS and krig-
ing versus the range of correlation of the underlying grade
models. The revenue attained by both methods increases as
the range increases. A very low range appears like a high
nugget effect, that is, negative revenue. The MPS proce- 1000. maximum revenue
dure leads to significantly greater revenue when the range MPS
of correlation is greater than the twice the blasthole spac-
ing.
800. kriging-based
The previous numerical experiments considered a small,
yet realistic sampling error. Pitard describes many sources
of errors in sampling. These errors can be significant, par-
600.
ticularly for gold deposits. Different levels of sampling er-
Revenue $k
we = z +
z error !, N (0; r
e z ) (6)
600.
Revenue $k
1.0
kriging-based
400.
0.0
0.
0.0 1.0 2.0 3.0
?
1.176 0.8944 1.042 1.392
10
Frequency
0.04
Figure 17: Blasthole layout for small scale example.
0.03
0.02
pattern. Note that different blasthole grades will be consid-
0.01
ered although the configuration of the blastholes and the
block will remain unchanged. 0.0
One hundred different realizations of the 16 blasthole 1.15 1.20 1.25 1.30 1.35
(2) kriging and MPS both classify the block as waste - 54 0.04
times, (3) kriging classifies the block as ore but MPS clas-
0.03
sifies it as waste - 4 times, and (4) kriging classifies the
block as waste and MPS classifies it as ore - 2 times. Al- 0.02
though most classifications are the same the 6 differences
0.01
are interesting.
Figure 18 shows the uncertainty in block grades at the 0.0
two locations where the kriged value would indicate waste 1.15 1.20 1.25 1.30 1.35
Realization 18 KO-SW
Number of Data 200
1.084 1.495 1.166 1.162
mean 1.2349 20
std. dev. 0.0150
0.12 coef. of var 0.0121
maximum
upper quartile
median
lower quartile
1.2823
1.2441
1.2340
1.2231
?
1.140 1.127 1.266 1.187
minimum 1.2045 10
Frequency
0.08
0.0
1.15 1.20 1.25 1.30 1.35
0 10 20 30
Realization 29 KO-SW
Number of Data 200 Configuration 29
mean 1.2181
std. dev. 0.0324
0.08 coef. of var 0.0266
maximum 1.3405
upper quartile 1.2380
median 1.2150
0.06 lower quartile 1.1940 1.522 1.647 1.144 1.089
minimum 1.1602 30
Frequency
0.04
0.0
?
1.209 1.114 1.368 1.124
1.15 1.20 1.25 1.30 1.35 10
Figure 19: Uncertainty in block grades at two locations where 1.044 0.8700 1.212 1.098
the kriged value would indicate ore (black dot) and simulation 0
(MPS) would indicate waste.
0 10 20 30
0.15
ogram (zero nugget effect and range 10 times the block
size). The importance of sampling errors will likely de-
0.10
pend on the underlying level of spatial continuity. This is
an area for further research / application.
0.05
Blast movement studies are necessary to quantify and
understand how the dig limits should be established af-
0.0 ter blasting. Improvements in grade control methodology
1.0 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8
could be reduced if there is significant throw (1 to 3 m of
throw could be expected). This may be acceptable pro-
vided the throw is in a single known direction; however, a
Figure 21: Example showing “blasthole” and “block” grades for general mixing would be difficult to overcome. This com-
central 5m by 5m block. The data from realization 29 were used.
Note that the mean is unchanged and the classical reduction in
plicating factor would have to be handled on a case-by-case
variance from blasthole grades to block grades. basis.
In general, there is more information for grade control
than the blasthole grades on the current bench. There are
center of the small grid would be significantly higher than geological rock types, exploration holes, blastholes from
the uncertainty in a “block” grade. In fact, these two dif- the benches above. The additional value of this information
ferent levels of uncertainty for realization 29 are shown on could be assessed by calculating the additional revenue us-
Figure 21. The block grades have less variability, in par- ing this data. Of course, it would depend on the closeness
ticular, less probability of being above the cutoff grade of of the data.
1.25. Computer software implementing the numerical algo-
From this small example we can draw two tentative rithms presented in this paper is not difficult, in fact, the
conclusions (1) when the block is borderline ore / waste authors would send FORTRAN code on request. There is a
with large uncertainty, the MPS procedure would tend to need, however, for the entire procedure to be automated in
classify it as ore, and (2) when the block is borderline but easy-to-use software. A familiar Windows-type interface
based on an isolated high grade sample, the MPS procedure would be logical.
is less sensitive to lone high grade values and would tend
to classify the block as waste. As seen in the preceeding
example, these effects lead to greater revenue / profit. CONCLUSIONS
Classifying material as ore and waste based on a maxi-
mum profit criteria generates more revenue than conven-
OUTSTANDING ISSUES tional kriging. The concepts used in this paper are well
Real data could have been used for the example, but it is established, that is, geostatistical simulation and basic eco-
very difficult to check the efficacy of the algorithm since nomics. Implementation is straightforward and the selec-
the truth is inaccessible. In other words, application of the tion program may be run on low-level Pentium PCs avail-
proposed methodology to real data is straightforward but able at virtually every mine site. The method is flexible in
its performance would be difficult to quantify. An impor- its ability to handle variability of costs, variability of metal
tant extension of this work is to present documented exam- recovery, and revenue from multiple ore minerals.
ples at operating mines. The method increases revenue by rigorous accounting
12 Copyright (c) 1999 by SME
of uncertainty in block grades. The effect of isolated high
grade values is minimized, the lost opportunity cost of wasted
ore is taken into consideration, and the irreducible uncer-
tainty is handled by considering the expected value of the
profit if called ore Pore and waste Pwaste .
Additional effort is required to apply at operating mines
with easy-to-use software and that can be straightforwardly
applied by the engineer / geologist / technician at the mine
site.
ACKNOWLEDGEMENTS
The financial support of the Natural Sciences and Engi-
neering Research Council of Canada (NSERC) is grate-
fully acknowledged.
REFERENCES
1. David, M., Geostatistical Ore Reserve Estimation,
Amsterdam, Elsevier, (1977).
2. Davis, B. M., Trimble, J. and McClure, D., Grade
Control and Ore Selection Practices at the Colos-
seum Gold Mine, Mining Engineering, August 1989,
pages 827-830.
3. Deutsch, C.V. and Journel, A.G.:GSLIB: Geostatis-
tical Software Library and User’s Guide, New York,
Oxford University Press, (1992), page 340.
4. Glacken, I. M., Change of Support by Direct Condi-
tional Block Simulation, Geostatistics Wollongong
’96, Kluwer Academic Publishers, vol. 1, September
1996.
5. Journel, A. G. and Huijbregts, Ch. J., Mining Geo-
statistics, New York, Academic Press, (1978), page
600.
6. Pitard, F. F., Pierre Gy’s Sampling Theory and Sam-
pling Practice,” New York, CRC Press, (1993), page
488.
7. Raymond, G. F., Ore Estimation Problems in an Er-
ratically Mineralized Orebody, CIM Bulletin, June
1979, pages 90-98.
8. Snowden, V., C. Moore and D. Kelly, Using Geo-
statistics to Assist in Optimising Grade Control Esti-
mation at KCGM’s Flimston Open Pit, Proceedings
of the 4th Large Open Pit Mining Conference, Perth,
Australia, Australasian Institute of Mining & Metal-
lurgy, September, 1994, pages 117-131.
9. Srivastava, R.M.: Minimum Variance or Maximum
Profitability?, CIM Bulletin, vol. 80, no. 901, 1987,
pages 63-68.
10. Westley, J. F. H., Product Monitoring, Grade Con-
trol, and Ore:Waste Reconciliation at Bougainville
Copper Limited (BCL), Proceedings of the Large
Open Pit Mining Conference, Newman, WA Aus-
tralia, Australasian Institute of Mining & Metallurgy,
October, 1986, pages 277-284.