Computers & Geosciences 51 (2013) 238–246

Contents lists available at SciVerse ScienceDirect

Computers & Geosciences

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

Ultimate open pit stochastic optimization

Denis Marcotte n, Josiane Caron
Department of Civil, Geological and Mining Engineering, Polytechnique de Montréal, C.P. 6079 Succ. Centre-ville, Montréal, QC, Canada H3C 3A7

a r t i c l e i n f o a b s t r a c t

Article history: Classical open pit optimization (maximum closure problem) is made on block estimates, without
Received 3 May 2012 directly considering the block grades uncertainty. We propose an alternative approach of stochastic
Received in revised form optimization. The stochastic optimization is taken as the optimal pit computed on the block expected
8 August 2012
profits, rather than expected grades, computed from a series of conditional simulations. The stochastic
Accepted 9 August 2012
Available online 18 August 2012
optimization generates, by construction, larger ore and waste tonnages than the classical optimization.
Contrary to the classical approach, the stochastic optimization is conditionally unbiased for the realized
Keywords: profit given the predicted profit. A series of simulated deposits with different variograms are used to
Open pit optimization compare the stochastic approach, the classical approach and the simulated approach that maximizes
Stochastic optimization
expected profit among simulated designs. Profits obtained with the stochastic optimization are
Geostatistical simulation
generally larger than the classical or simulated pit. The main factor controlling the relative gain of
stochastic optimization compared to classical approach and simulated pit is shown to be the
information level as measured by the boreholes spacing/range ratio. The relative gains of the stochastic
approach over the classical approach increase with the treatment costs but decrease with mining costs.
The relative gains of the stochastic approach over the simulated pit approach increase both with the
treatment and mining costs. At early stages of an open pit project, when uncertainty is large, the
stochastic optimization approach appears preferable to the classical approach or the simulated pit
approach for fair comparison of the values of alternative projects and for the initial design and planning
of the open pit.
& 2012 Elsevier Ltd. All rights reserved.

1. Introduction possible to define, by stochastic optimization, more profitable

open pit limits than with the simple deterministic approach based
The definition of optimal pit limits and profit is an essential on estimates (usually obtained by ordinary kriging) (Henry et al.,
part in pre-feasibility and feasibility studies on open pit mines. 2005; Bastante et al., 2008; Meagher et al., 2009). The aim is to
The pit limits define the ore and waste tonnages and the ore determine under which conditions deterministic and stochastic
values. It is the basis for important initial decisions such as the solutions are almost equivalent and under which conditions do
production capacity, the choice of equipments, and the location they differ significantly. We want to quantify this difference. Only
and size of surface installations (crushing plan, stockpiles, waste the geological uncertainty is considered in this study.
dump, offices, etc.). Moreover, the project internal rate of return The effect of grade uncertainty on value of open pit and design
can be computed and compared with other investment opportu- was studied by Dowd and Onur (1993), Dowd (1994), Whittle and
nities. The pit limits of the optimal pit maximize the net present Bozorgebrahimi (2004), and Froyland et al. (2004). Froyland et al.
value (NPV). The classical problem of pit optimization is easily (2004) suggested to optimize the block expected NPV computed
solvable using well known and efficient algorithms like the Lerchs from a series of conditional simulations. The effect of grade un-
and Grossmann (1965) and the max-flow (Picard, 1976). An ex- certainty in open pit operations like pushback design and sche-
cellent review of different implementations and algorithms for duling problems was studied by Ramazan and Dimitrakopoulos
open pit optimization is found in Hochbaum (2001). (2007), Bastante et al. (2008), Meagher et al. (2009), and
In practice, pit optimization is performed on blocks which true Armstrong and Galli (2011). Dimitrakopoulos and Abdel Sabour
grades are unknown and can only be estimated or simulated (2007) studied the effect of uncertainty for long-term planning for
using the available information. Also, future commodity prices an Australian mine. They report 11% higher money value for the
are uncertain. Because of these uncertainties, it is theoretically design incorporating uncertainty. Henry et al. (2005) studied the
effect of both grade and price uncertainty on a small deposit
and found an increase value of 6%. Are these results transposable
n
Corresponding author. to most mines? In fact, no study was devoted to quantify the
E-mail address: [email protected] (D. Marcotte). gain expected with the stochastic optimization compared to

0098-3004/$ - see front matter & 2012 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.cageo.2012.08.008
 D. Marcotte, J. Caron / Computers & Geosciences 51 (2013) 238–246 239

deterministic optimization over a wide variety of (synthetic) When the true block grade is uncertain, it is tempting to apply
deposits and under varying mining and treatment costs. The main directly the profit function to the (unbiased) estimated block
objective of this paper is to fill this gap. grade. However, due to the convexity of the profit function (see
We use geostatistical simulations to (i) generate synthetic Eq. (1)) this approach would underestimate the expected profit of
deposits with different spatial characteristics, (ii) obtain samples the block. Indeed, a block can receive a different assignment when
from the synthetic deposits in pseudo-boreholes, and (iii) produce its grade becomes perfectly known. The new assignment based on
conditional simulations to the borehole grades. The conditional the true grade can only increase or leave unchanged the value of
simulations are used for stochastic optimization based on the the block compared to the value recovered for that block when
expected NPV of each block. The effects of the grade spatial the assignment is not updated (dashed straight line in Fig. 1). The
continuity characteristics and of the mining and treatment costs present expected value can be computed when the distribution of
on the predicted and realized NPV for the different synthetic block grade estimation error is known. For illustration, Fig. 1
deposits are studied, both for the stochastic and deterministic (right) shows the expected profit of the block as a function of
approaches. Moreover, our proposed stochastic method is com- estimation error standard deviation for a Gaussian distribution.
pared to the approach of Dimitrakopoulos et al. (2007a) based on Clearly, the expected profit of a block is always equal or greater
the designs from conditional simulations. We stress that our than the profit computed with the estimated grade. The differ-
study focus simply on the ultimate open-pit design, not on the ence increases with the standard deviation and for a given
more complex scheduling or short- to mid-term mining operation standard deviation, the difference is maximum when the product
problems. g  r is close to the cut-off grade (treatment costs) where future
changes in assignments are more likely to occur.
A direct consequence of Fig. 1 is that the optimal pit based on
block expected profits (psto) is always larger than the optimal pit
2. Methods
based on the block estimates (pest). Moreover, the predicted profit
of the optimal pit based on psto is always larger than the one
2.1. Profit function
predicted with pest. This does not imply that the profit realized
within psto pit limits is necessarily larger than within pest pit limits
In a mine, real block grades are never exactly known. The
(although, in our simulations, it is generally the case). Moreover,
uncertainty is even larger at the time pit optimization is done for
psto is unbiased by construction while pest predicted pit profit
feasibility studies. The block grades must be estimated/simulated
underestimates the real profit available within the pit limits.
using available data. However, at the time of mining a given
block, more information will be available (e.g. from blast holes
analysis), and the block grade will be eventually well known. 2.2. Open pit optimization
Therefore, the final destination of a block, based on this new
knowledge, can be different from its initial assignment. To The open pit optimization maximizes the pit profit under slope
simplify the analysis, and without loss of generality, we assume: constraints represented by precedence relations between blocks
(i) that only two destinations for blocks within the pit are possible (Fig. 2). For simplicity, we use fixed metal price and costs and
(either the block is treated at the mill or it is send to waste apply no discount rate to the block profit.
dump); (ii) the cost of mining a waste block is assumed equal to The linear programming model is then
the cost of mining an ore block and (iii) the cost of mining is X
N

assumed independent of depth. Under these assumptions, all Maximize pi xi ð2Þ

i¼1
blocks with grade covering the treatment costs should be treated.
The profit function associated to any block grade is then as Subject to xi xj r 0, i A V, j A Gi ð3Þ
depicted in Fig. 1 (left). The mining and treatment costs are
expressed, for convenience, as grades. The profit p per unit xi ¼ 0 or 1, 8iA V ð4Þ
tonnage of a block having grade g, with recovery factor r, mining
where pi, i ¼ 1, . . . ,N is the profit assigned to block i, V is the set of
costs cm and treatment costs ctr is then given as
blocks, Gi is the subset of blocks that must be mined to gain
access to block i (precedence relations). This problem is known as
p ¼ g  rcm ctr when g  r Zctr
the maximum closure problem. Picard (1976) was the first one to
p ¼ cm when g  r octr ð1Þ show how the maximum closure problem can be solved using

s1
s2
s3

Waste Ore
Net value

Net value

0
m.c. m.c.

m.c.+ t.c. m.c.+ t.c.

t.c. t.c.
Grade Grade

Fig. 1. Left: profit function for a block with known grade. Right: expected block profit as a function of the block grade expectation and standard deviation, with s1 o s2 o s3 .
240 D. Marcotte, J. Caron / Computers & Geosciences 51 (2013) 238–246

efficient max-flow algorithms. Hochbaum (2001) confirmed that factor of 0.9. Also, a second block profit measure is obtained by
best implementations of max-flow algorithm are significantly first converting each conditional simulation to a profit pisto using Eq.
more efficient than the classical Lerchs and Grossman’s algorithm. (1), with a recovery factor of 0.9, and then averaging the 30 profits.
The optimization on psto represents the stochastic optimization
2.3. Simulated deposits and grade uncertainty whereas the optimization on pest is the deterministic optimization
classically found in pre-feasibility and feasibility studies.
A series of 12 different spatial structures (spherical covariance Open pit optimization is done on both pest and psto for a series of
with different ranges and with/without nugget effect) are simu- different mining costs (0.10, 0.20, 0.30 and 0.40) and treatment costs
lated (see Table 1). All deposits have a theoretical mean of 1. For (0.40, 0.60, 0.80, 1.0, 1.2). A crown of zero profit blocks is padded all
each structure, 20 deposits are generated. All deposits are around each simulated deposit so as to allow each simulated block
represented by nx ¼159  ny ¼159  nz ¼45 points, 3 m apart in to be eventually part of the optimal pit. Five precedence links are
each direction. All simulations are done in the Gaussian space and used (see Fig. 2). The predicted pit profits are computed for each
then transformed by exponentiation to give a lognormal point costs scenarios, each deposit, and the two block profit indexes pest
distribution. Cubic blocks (9 m  9 m  9 m) are formed by aver- and psto. The realized profits are also computed by applying the
aging 27 points per block, for a total of 53  53  15¼ 42 135 optimal selections obtained with pest and psto to the real deposits. In
blocks per deposit. Boreholes are obtained as additional point all, 12  20  4  5¼4800 optimal pits are computed for (i) the ‘real’
values (not used in the block grade computation), each point deposit (ideal pits), (ii) the pest and (iii) the psto. An additional
value representing 3 m length in the borehole. A total of 8  8¼64 4800  30¼144,000 optimal pits are computed with the conditional
boreholes on a regular 60 m centered grid, for a total of 2880 data, is simulations. These ‘conditional pits’ are used solely to evaluate the
used for all deposits. For each deposit, a series of 30 conditional (to potential of using one of the simulated pits as best design pit as
the borehole grades) simulations are obtained and block conditional proposed by Dimitrakopoulos et al. (2007a).
grades are computed the same way as for the deposit. This provides Due to the size of the simulated field and the large number of
the block grade distribution, compatible with the borehole data and pits to optimize, an efficient pit optimizer is indeed required.
deposit spatial statistics, required for stochastic optimization. Optimization is done using max-flow push-relabel algorithm of
For each simulated deposit, block grade estimates are obtained Goldberg and Tarjan (1986) available within the Matlab toolbox
by averaging the 30 conditional simulations. This represents a matlab bgl (Gleich, 2007). All the geostatistical simulations are
smooth grade estimate similar to kriging. The estimated block done using the turning bands method and post-conditioning by
grades are converted into profits pest using Eq. (1) with a recovery kriging (Chile s and Delfiner, 1999).

Table 1
Spherical covariance model parameters.

Deposit # C0 C Range (m)

1 0 4 60
2 0 4 120
3 0 4 180
4 0 1 60
5 0 1 120
6 0 1 180
7 2 4 60
8 2 4 120
9 2 4 180
10 0.5 1 60
11 0.5 1 120
12 0.5 1 180
Fig. 2. Example of precedence relations between blocks in an open pit mine.

x 104 Predicted pit profit x 104 Realized pit profit

5 5

4 4

3 3
Profit with psto

Profit with psto

2 2

1 1

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Profit with pest x 104 Profit with pest x 104

Fig. 3. Left: predicted (left) and realized (right) pit profits after optimization on pest and psto for the 240 simulated deposits and 20 costs scenarios (total 4800 points).
 D. Marcotte, J. Caron / Computers & Geosciences 51 (2013) 238–246 241

3. Results 3.2. Factors controlling the gain obtained by psto over pest

3.1. Predicted and realized profits Fig. 7 shows the influence of the model parameters on the
gains (expressed as %) obtained by psto over pest. Clearly, the
As indicated previously, because of the convexity of the profit dominant factor for the gain expected by stochastic optimization
function, one has: pisto Z piest ,8i ¼ 1 . . . N, where N¼ 30 is the number of is the range (or correlation distance) of the variogram model. On
conditional simulations for a given deposit. Consequently, as Fig. 3 the opposite, the nugget effect and the variance do not have a big
shows the pit predicted profit with psto is always larger than with pest. impact on the observed relative gains. The biggest gains are
Also, the realized pit profits with psto are generally higher than with pest. obtained when the range is equal to the borehole spacing. When
In average, pits optimized on psto produce 2.7% more profits than the information is denser relative to the range (ranges 120 m and
the pits optimized on pest. As Fig. 4 shows, the lower is the realized 180 m), the gain is generally small. In these cases, both psto and
profit with pest, the bigger is the relative gain obtained with psto. For pest pits are closer to the ideal one, hence their differences tend to
very rich open pits, optimization on psto or pest barely matters in vanish. This interpretation is confirmed by Fig. 8. Hence, gains
relative terms although in absolute value it might be non-negligible. obtained by psto are expected to be larger at the early stages of
open pit mine planning when the data is scarce.
3.1.1. Bias Fig. 9 shows the influence of the costs on the gains (expressed as
Fig. 5 shows the realized and expected profits for optimization %) obtained by psto over pest. The gains increase with the treatment
on psto and pest. As expected, a significant underestimation occurs costs but tend to diminish with the mining costs. As the treatment
for the pit profits based on pest (10.0% underestimation) whereas costs increase, it gets closer to the deposit mean grade where the
psto is approximately unbiased (1.0% overestimation). For pest, the non-linearity of the profit function has the stronger effect on the
bias is generally larger for poorer deposits (Fig. 6), whereas it is expected profit (see Fig. 1). On the other hand, when the mining
everywhere close to 0 for psto. costs increases, the pits get smaller and there is less room to realize
gains by stochastic optimization over the classical approach.
25

40

20 30
Bias %
Gain (%) (psto−pest)/pest

20

15
10

0
0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5 2.7 2.9
10 Realized profit with pest
20

5 10
Bias %

0 −10
0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5 2.7 2.9
Realized profit with pest x 104 −20
0.1 0.3 0.5 0.7 0.9 1.1 1.3 1.5 1.7 1.9 2.1 2.3 2.5 2.7 2.9
Fig. 4. Relative mean gain in realized pit profit after optimization on psto compared Realized profit with psto
to realized profit with pest as a function of realized profit with pest. The mean gain is
computed over all pits showing a realized profit with pest that lies within 7 1000 Fig. 6. Predicted and realized profits for pest (top) and psto (bottom) for the 240
of the value indicated on the abscissa for each bar. simulated deposits and 20 costs scenarios.

x 104 pest x 104 psto

5 5

4 4

3 3
Realized

Realized

2 2

1 1

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Predicted x 104 Predicted x 104

Fig. 5. Predicted and realized profits for pest (left) and psto (right) for the 240 simulated deposits and 20 costs scenarios.
242 D. Marcotte, J. Caron / Computers & Geosciences 51 (2013) 238–246

3.3. Factors controlling the profits realized by psto mean penalty incurred for not knowing perfectly the orebody at
the time of optimization. The penalty is relatively larger for
Fig. 10 shows the influence of the model parameters on the marginal pits than for rich pits.
profits realized by psto. The profit increases with the value of the
variance parameter C. The effect of the range is unclear as the
3.5. Profits realized by pits based on conditional simulations
profits are sometimes maximum for the intermediate range 120.
The impact of nugget effects is limited.
It has been proposed in Dimitrakopoulos et al. (2007b) to
Fig. 11 shows the influence of the costs on the profits realized
select one of the pits optimized on a series of conditional
by psto. As expected, the profits realized with psto diminish both
simulations as a possibly better pit design. Suppose N conditional
with the treatment costs and the mining costs.
simulations are obtained. For each realization, an optimal pit is
computed. For realization i, the expected realized profit of the pit
3.4. Comparison of stochastic and ideal pits can be estimated by applying the design to the N1 other
realizations (note that the realization i should not be used in this
The ideal pit is the pit obtained by optimization on the profits computation to avoid overestimate the expected realized profit).
computed with the true block grade, i.e. assuming perfect knowl- The pit among the N designs that maximizes the expected
edge. The ideal pit realizes, by definition, more profit than any realized profit, or/and other suitable operational criteria, is
other pit (Froyland et al., 2004). Fig. 12 shows the ideal pit profit selected.
against the optimal pit based on psto. The mean gain of the ideal Fig. 13 (left) shows that the above strategy gives almost surely
pit profit over the psto realized profit is 6.6%. This represents the (97.5% of the times) worst results than the pit based on psto. The

50 50

40 40
Gain psto over pest

30 30
Gain psto over pest

20
20

10
10

0
0

−10
60 60 60 60 120 120 120 120 180 180 180 180 Range −10
0 0 0.5 2 0 0 0.5 2 0 0 0.5 2 Nugget 0.4 0.4 0.4 0.4 0.6 0.6 0.6 0.6 0.8 0.8 0.8 0.8 1 1 1 1 1.2 1.2 1.2 1.2 tr.c.
1 4 1 4 1 4 1 4 1 4 1 4 C 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 m.c.

Fig. 7. Boxplots of realized gains by psto over pest (in %) as a function of variogram Fig. 9. Boxplots of realized gains by psto over pest (in %) as a function of mining
model parameters. Box horizontal bars are traced at 25, 50 and 75th percentiles of (m.c.) and treatment costs (tr.c.). Box horizontal bars are traced at 25, 50 and 75th
the distribution (n¼ 400). Dashed lines (whiskers) are 1.5 times the box length on percentiles of the distribution (n¼ 240). Dashed lines (whiskers) are 1.5 times the box
both sides, points falling outside the whiskers are plotted as þ marks. length on both sides, points falling outside the whiskers are plotted as þ marks.

100
Gain over pest

80

60

40

20

0
60 60 60 60 120 120 120 120 180 180 180 180 Range
0 0 0.5 2 0 0 0.5 2 0 0 0.5 2 Nugget
1 4 1 4 1 4 1 4 1 4 1 4 C

100
Gain over psto

80

60

40

20

0
60 60 60 60 120 120 120 120 180 180 180 180 Range
0 0 0.5 2 0 0 0.5 2 0 0 0.5 2 Nugget
1 4 1 4 1 4 1 4 1 4 1 4 C

Fig. 8. Boxplots of realized gains by pideal over pest (top) and psto (bottom) (in %) as a function of variogram model parameters. Box horizontal bars are traced at 25, 50 and
75th percentiles of the distribution (n ¼400). Dashed lines (whiskers) are 1.5 times the box length on both sides, points falling outside the whiskers are plotted as þ marks.
 D. Marcotte, J. Caron / Computers & Geosciences 51 (2013) 238–246 243

x 10
3 x 104
5

2.5

2 4
Profit psto

1.5

Profit with psto

3
1

0.5
2

0
60 60 60 60 120 120 120 120 180 180 180 180 Range
0 0 0.5 2 0 0 0.5 2 0 0 0.5 2 Nugget
1 4 1 4 1 4 1 4 1 4 1 4 C 1

Fig. 10. Boxplots of profits realized by psto as a function of variogram model

parameters. Box horizontal bars are traced at 25, 50 and 75th percentiles of the
distribution (n¼ 400). Dashed lines (whiskers) are 1.5 times the box length on
0
both sides, points falling outside the whiskers are plotted as þ marks.
0 1 2 3 4 5
Ideal pit profit x 104
3
Fig. 12. Realized pit profit after optimization on psto against ideal pit profit for the
12  20 deposits and the 4  5 costs scenarios (total 4800 points).
2.5

with psto are larger than with pest. The average tonnage obtained by
2 psto over the 4800 pits is 27% larger than with pest. This percentage
climbs to 51% when considering only the pits obtained on models
Profit psto

with range 60 m (i.e. equal to the borehole spacing). Hence, the

1.5
greater the uncertainty on block grades, the larger is the increase in
pit size due to inclusion of uncertainty in the optimization.
1 Fig. 17 (right) shows the tonnage (in number of blocks) of the
4800 pits optimized on psto compared to the ideal pits tonnage.
The average tonnage obtained by psto over the 4800 pits is 12%
0.5
larger than the tonnage average for the ideal pits, the percentage
being larger with larger pits.
0 Finally, compared to the best simulated pits, the tonnage is
0.4 0.4 0.4 0.4 0.6 0.6 0.6 0.6 0.8 0.8 0.8 0.8 1 1 1 1 1.2 1.2 1.2 1.2 tr.c.
0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 m.c. only 9.7% larger with psto. The tonnage difference is negatively
correlated to the range.
Fig. 11. Boxplots of profits realized by psto as a function of mining (m.c.) and
treatment costs (tr.c.). Box horizontal bars are traced at 25, 50 and 75th percentiles
of the distribution (n¼ 240). Dashed lines (whiskers) are 1.5 times the box length on
both sides, points falling outside the whiskers are plotted as þ marks. 4. Discussion

mean gain in profits for psto over the best simulated pit is 4.1%, and For a wide variety of deposits and costs structure, the open pits
the gain of psto is larger with poorer deposits (see Fig. 14). The optimized on psto have been shown to be significantly more
relative gains obtained with psto over the best simulated pit increase profitable than those obtained on pest, the latter corresponding
with mining and treatment costs (see Fig. 15) and are larger when to the usual practice. The psto optimized pits are also significantly
the range is equal to the borehole spacing (see Fig. 16). more profitable than the best pits obtained on conditional simula-
In fact, the same observations hold when comparing the tions. The relative gains of psto is maximum for poorer deposits
profits obtained by psto this time with the pit design that provides where it is of the order of 10–25%. For richer deposits, the gains of
the maximum profit on the true deposit (note that in practice the psto decrease as the pit tends to become richer with all approaches.
identification of this design is impossible as the true deposit is The psto approach assumes the final selection of a given block,
imperfectly known). Fig. 13 (right) shows the scattergram of the at the time of mining the block, will be based on an improved
realized profits with psto and the maximum realized profit on the estimate very close to the real grade. The pest approach on the
real deposit among the N ¼30 pits of the conditional simulations contrary assumes that the block grade knowledge does not
for each of the 12  20 deposits and 4  5 costs scenarios. The best improve so the final selection remains based on the initial
pit from conditional simulations is 85% of the times less profitable estimate. Neither assumption is totally valid as, on the one hand,
than the pit based on psto. On average, pit profits generated with perfect knowledge is out of reach and, on the other hand, new
psto are 1.9% higher than the profits obtained with the best pit of data acquired during mine’s life will necessarily improve preci-
the conditional simulations. sion of future estimates. Therefore, in real operations, results will
lie somewhere between those of psto and pest. A first consequence
3.6. Tonnage of open pits is that the underestimation bias described for pest will be reduced
and psto is likely to present a small overestimation bias. A second
Fig. 17 (left) shows the tonnage (in number of blocks) of the consequence is that the relative gain obtained by using the
4800 pits optimized on psto and pest. As mentioned, all the pits stochastic approach is likely to be less than expected.
244 D. Marcotte, J. Caron / Computers & Geosciences 51 (2013) 238–246

x 104 x 104
5 5

4 4

Profit with cond. sim.

3 3

Max(sim)
2 2

1 1

0 0
0 1 2 3 4 5 0 1 2 3 4 5
Profit with psto x 104 Profit with psto x 104

Fig. 13. Left: realized pit profit after optimization on psto vs realized profit with the pit, among the 30 conditional simulations, providing maximum profit on the 29 other
conditional simulations. Right: realized pit profit after optimization on psto vs realized profit with the pit, among the 30 conditional simulations, providing maximum profit
on the true deposit. One point for each of the 12  20 deposits and the 4  5 costs scenarios (total 4800 points).

30 50

40
25 Gain of psto over best sim.

30
20
Gain of psto %

20

15

10

10
0

5
−10
0.4 0.4 0.4 0.4 0.6 0.6 0.6 0.6 0.8 0.8 0.8 0.8 1 1 1 1 1.2 1.2 1.2 1.2 tr.c.
0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 0.1 0.2 0.3 0.4 m.c.

0
0 0.5 1 1.5 2 2.5 3 Fig. 15. Boxplots of realized gains (in %) by psto, over the best simulated pit, as a
Profit of best simulated pit function of mining (m.c.) and treatment (tr.c.) costs. Box horizontal bars are traced
at 25, 50 and 75th percentiles of the distribution (n ¼240). Dashed lines (whiskers)
Fig. 14. Gain on realized profit of psto over the pit among the 30 conditional are 1.5 times the box length on both sides, points falling outside the whiskers are
realizations providing maximum profit on the 29 other conditional realizations. plotted as þ marks.

The level of information, in relation to the grade spatial pit design providing the largest expected realized profit (mean profit
continuity, was found to be an important factor controlling the obtained by design i on the N1 other realizations).
gain realized by using psto rather than pest. When the boreholes The profit function was not discounted. This has no important
spacing is equal to the correlation range, maximum gains are effect on the results of our study providing the discount rate
observed for psto. For closer boreholes relatively to the range, the remains always positive. In effect, it is well known that the open
gain expected is much lower (see Fig. 10). Under international pit problem can be parametrized (Francois-Bongarcon and Guibal,
regulations (e.g. NI-43-101), it is debatable whether or not 1982; Coleou, 1989). By varying the parameter value, a series of
estimation based on boreholes spaced at correlation length could nested pits is created. The nested pits have the remarkable
qualify the blocks as measured or indicated resources (proven or property that the mean predicted value per ton, between two
probable reserves). Hence, for feasibility studies based on proven successive nested pits, is necessarily decreasing from the smaller
reserves, the difference between psto and pest is likely to be to the larger pit (Picard and Smith, 1994). Hence the final pit,
modest. Thus, a large difference obtained between the stochastic discounted or not, remains the same as the profit of the blocks
optimization and the classical optimization is indicative of large contained within the two last parametric contour limits remains
uncertainty about the deposit and is a clear indication that further positive after discounting.
sampling could be useful for design purpose. On the other hand, In this study, metal price scenarios were considered fixed.
at early stages of the mine assessment, results based on psto could However, mining costs and treatment costs were both varied
justify to keep projects alive that would be otherwise abandoned. independently over a wide range of values (by a factor of 1–4 for
Our results indicate that it is better to select the pit optimized mining costs and a factor of 1–3 for treatment costs). As the main
on psto than any of the pits obtained on conditional simulations. findings on psto were valid for each set of mining and treatment
Mean gain of 1.9% and 4.1% was obtained in favor of psto compared costs (see Figs. 9 and 11), it can be safely assumed that similar
to, respectively, the simulated pit design showing the largest realized results would have been obtained with metal prices varying
profit on the true deposit (in practice inaccessible) and the simulated over time.
 D. Marcotte, J. Caron / Computers & Geosciences 51 (2013) 238–246 245

50

40

Gain of psto over best sim.

30

20

10

−10
60 60 60 60 120 120 120 120 180 180 180 180 Range
0 0 0.5 2 0 0 0.5 2 0 0 0.5 2 Nugget
1 4 1 4 1 4 1 4 1 4 1 4 C

Fig. 16. Boxplots of realized gains (in %) by psto over the best simulated pit as a function of variogram model parameters. Box horizontal bars are traced at 25, 50
and 75th percentiles of the distribution (n ¼ 400). Dashed lines (whiskers) are 1.5 times the box length on both sides, points falling outside the whiskers are plotted
as þ marks.

x 104 x 104
7 7

6 6
Tonnage with psto (nb. of blocks)
Tonnage with psto (nb. of blocks)

5 5

4 4

3 3

2 2

1 1

0 0
0 1 2 3 4 5 6 7 0 1 2 3 4 5 6 7
Tonnage with pest (nb. of blocks) x 104 Tonnage with ideal pits (nb. of blocks) x 104

Fig. 17. Tonnage with psto vs tonnage with pest (left) and tonnage with ideal pit (right) for the 240 deposits and 20 costs scenarios (total 4800 points).

5. Conclusion with the mining and treatment costs. Finally, it was shown that the
stochastic approach is unbiased, contrary to the classical approach.
The proposed approach of stochastic optimization is based on We conclude that the stochastic optimization approach presented is
the expected profit of each block computed from a series of a better tool than the classical approach or best simulated pit for fair
conditional simulations. The stochastic optimization was pro- comparison of the values of alternative projects and for the initial
ven to always generate larger ore tonnages and generally design and planning of the open pit.
larger profits than the optimization based on estimated grades,
the classical approach. It was also shown to be almost invari-
ably more profitable than the best simulated pit (the pit with Acknowledgments
largest expected realized profit). The main factor controlling
the relative gain of stochastic optimization compared to This research was made possible by financial support from
classical optimization or best simulated pit was shown to be NSERC of Canada, FQRNT of Québec and the contribution of
the information level as measured by the boreholes spacing/ Corporation minie re Osisko in the form of an industrial scholar-
range ratio. The gains compared to classical approach increase ship. Helpful discussions with C. Audet and contribution of M.
with the treatment costs but decrease with mining costs. Gamache who brought to our attention efficient max-flow algo-
Compared to the best simulated pit, the gains increased both rithm implementations are also acknowledged. Two anonymous
246 D. Marcotte, J. Caron / Computers & Geosciences 51 (2013) 238–246

reviewers provided constructive comments that were helpful to Francois-Bongarcon, D., Guibal, D., 1982. Algorithms for parameterizing reserves
improve the paper. under different geometrical constraints. In: 17th International APCOM Sym-
posium, Golden, CO, pp. 297–309.
Froyland, G., Menabde, M., Stone, P., Hodson, D., 2004. The value of additional
References drilling to open pit mining projects. In: Proceedings of the Orebody Modelling
and Strategic Mine Planning—Uncertainty and Risk Management, Perth,
pp. 169–176.
Armstrong, M., Galli, A., 2011. A new approach to flexible open-pit optimisation Gleich, D., 2007. Matlab BGL v2.1 /http://www.mathworks.comS.
and scheduling. In: 35th APCOM Symposium, 2011. Goldberg, A.V., Tarjan, R.E., 1986. A new approach to the maximum flow problem.
Bastante, F.G., Taboada, J., Alejano, L., Alonso, E., 2008. Optimization tools and In: Proceedings of the Eighteenth Annual ACM Symposium on Theory of
simulation methods for designing and evaluating a mining operation. Sto- Computing. ACM, Berkeley, CA, United States, pp. 136–146.
chastic Environmental Research and Risk Assessment 22, 727–735. Henry, E., Marcotte, D., Samis, M., 2005. Valuing a mine as a portfolio of european
Chile s, J.-P., Delfiner, P., 1999. Geostatistics: Modeling Spatial Uncertainty. Wiley, call options: the effect of geological uncertainty and implications for strategic
New York. planning. In: Leuangthong, Deutsch, (Eds.), Geostatistics Banff 2004. Quanti-
Coleou, T., 1989. Technical parametrization of reserves for open pit design and tative Geology and Geostatistics. Springer, Banff, pp. 501–510.
mine planning. In: Weiss, A. (Ed.), 21st Application of Computers and Hochbaum, D.S., 2001. A new[space]–[space]old algorithm for minimum-cut and
Operations Research in the Mineral Industry Proceedings, pp. 485–494. maximum-flow in closure graphs. Networks 37, 171–193.
Dimitrakopoulos, R., Martinez, L., Ramazan, S., 2007a. A maximum upside/mini- Lerchs, H., Grossmann, I.F., 1965. Optimum design of open-pit mines. Canadian
mum downside approach to the traditional optimization of open pit mine Institute of Mining and Metallurgy 58, 47–54.
design. Journal of Mining Science 43, 73–82. Meagher, C., Abdel Sabour, S.A., Dimitrakopoulos, R., 2009. Pushback design of
Dimitrakopoulos, R., Martinez, L., Ramazan, S., 2007b. Optimising open pit design open pit mines under geological and market uncertainties. In: AusIMM (Ed.),
with simulated orebodies and whittle four-x—a maximum upside/minimum Orebody Modelling and Strategic Mine Planning, Perth, WA, pp. 297–304.
downside approach. In: Australasian Institute of Mining and Metallurgy Picard, J.-C., 1976. Maximal closure of a graph and applications to combinatorial
Publication Series. Perth, Australia, Australasian Institute of Mining and problems. Management Science 22, 1268–1272.
Metallurgy, pp. 201–206. Picard, J.C., Smith, B., 1994. Optimal Rate of Return in Open Pit Mine Design. École
Dimitrakopoulos, R.G., Abdel Sabour, S.A., 2007. Evaluating mine plans under Polytechnique.
uncertainty: Can the real options make a difference?. Resources Policy 32, Ramazan, S., Dimitrakopoulos, R., 2007. Stochastic optimisation of long-term
116–125. production scheduling for open pit mines with a new integer programming
Dowd, P., 1994. Risk assessment in reserve estimation and open-pit planning. formulation. In: Australasian Institute of Mining and Metallurgy Publication
Transactions of the American Institute of Mining and Metallurgical Engineers Series. Australasian Institute of Mining and Metallurgy, Perth, WA, Australia,
103, A148–A154. pp. 385–391.
Dowd, P., Onur, A., 1993. Open pit optimisation: I. Optimal open pit design. Whittle, D., Bozorgebrahimi, A., 2004. Hybrid pits—linking conditional simulation
Transactions of the American Institute of Mining and Metallurgical Engineers, and Lerchs–Grossmann through set theory. In: AusIMM (Ed.), Orebody
A94–A104. Modelling and Strategic Mine Planning, pp. 399–404.