Download as pdf or txt
Download as pdf or txt
You are on page 1of 25

Downloaded form http://iranpaper.

ir

Mathematical Geology, Vol. 37, No. 5, July 2005 (


C 2005)

DOI: 10.1007/s11004-005-6660-9

Grade Control Classification of Ore and Waste:


A Critical Review of Estimation and Simulation
Based Procedures1
Georges Verly2

For a mining operation, minimizing misclassification of ore and waste during grade control is of
critical importance. This paper reviews grade control procedures for classifying ore and waste that
are based on estimation and conditional simulation approaches. The similarities and differences of the
procedures are first discussed in the light of an example. Subsequently, several case studies, including
actual operating mines, are presented. The theoretical review shows that conditional simulation is
generally better than estimation for ore and waste classification, especially when the profit value of a
block is a nonlinear function of the grade of the block. The case studies, however, indicate that, in the
context of daily grade control, the value of conditional simulation over good estimation is relatively
low compared to the value of improving general practices at mine sites.

KEY WORDS: grade control, conditional simulation, nonlinear functions, sequential Gaussian
simulation.

INTRODUCTION

Minimizing the misclassification of ore and waste during grade control is important
as it directly increases the revenue of mining operations. In recent years, the use of
conditional simulation for improved grade control has been proposed and adopted
in mines, particularly in Australia (Collet and Corley, 2000; Deutsch and others,
1999; Glacken, 1997; Godoy and others, 2001; Isaaks, 1990; Schofield and Rolley,
1997; Shaw and Khosrowshari, 1997; Srivastava, 1987).
The simplest implementation of grade control consists of manually designing
ore-waste boundaries or ore blocks on a map of blast hole grade values. This
method is not very good because the map used to design ore blocks is equivalent to
a map of polygonal estimates. Better results are obtained when the map is estimated
by kriging or inverse square distance. The main argument for using simulation
is that smoothed maps obtained by kriging do not account for the uncertainty in

1Received 29 September 2002; accepted 14 May 2003.


2Placer Dome Inc., P.O. Box 49330, Bentall Station, Vancouver, British Columbia, Canada V7X 1P1;
e-mail: georges verly@placerdome.com.

451
0882-8121/05/0700-0451/1 
C 2005 International Association for Mathematical Geology
Downloaded form http://iranpaper.ir

452 Verly

the grade estimation or for the economic consequence of misclassification. As a


result, optimum classification cannot be achieved (Isaaks, 1990).
This paper reviews the two estimation-based grade control methods men-
tioned above, that is, the simple polygonal method (abbreviated as Poly); and the
ordinary kriging approach (abbreviated as OK). It also reviews four simulation-
based methods. One method is similar to ordinary kriging, but has average simu-
lated grades instead of kriged grades (abbreviated as AvgSim). The other methods
are the minimum loss approach (Isaaks, 1990; Srivastava, 1987) (abbreviated
as MinLoss), a maximum profit approach (abbreviated as MaxProf), and various
profit/loss approaches (Deutsch and others, 1999; Glacken, 1997) (abbreviated as
ProfLoss).
Ordinary kriging (David, 1977; Isaaks and Srivastava, 1989; Journel and
Huijbregts, 1978) and sequential Gaussian simulation (Gomez-Hernandez and
Journel, 1993; Goovaerts, 1997; Isaaks, 1990; Verly, 1993) will be used in this
paper. Other kriging or simulation methods could have been considered, for exam-
ple, indicator kriging or simulation and probability field simulation. However, the
purpose of the paper is not to report on the various kriging or simulation methods,
but rather to investigate the process that is followed and the results obtained.
In the next sections, the grade control methods are reviewed and critically
compared in a simple theoretical framework. Geometrical interpretations are given.
Results obtained on the Walker Lake data (Isaaks and Srivastava, 1989) and at two
operational mines are presented. Summary and conclusions follow.

REVIEW OF GRADE CONTROL METHODS

Nomenclature

Following the literature mentioned above, the metal grade is represented


herein by z and two categories for grade control are defined: ore and waste. Of
the economic parameters, p stands for metal price; r for metal recovery factor; cm
for mining cost per ton incurred for both ore and waste; and cp for ore-processing
cost per ton.
The marginal breakeven cut-off grade zc is defined as the grade at which the
ore revenue equals processing cost:
cp
przc − cm − cp = −cm ⇔ zc = . (1)
pr
Note that a higher cut-off grade could also be defined to account for addi-
tional fixed costs. This, however, does not change the fundamentals of the results
presented in this paper.
There are four possible scenarios for a block of grade z: (1) correct
acceptance—if the block grade is above the cut-off grade (z > zc ); (2) false
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 453

acceptance—if the block grade is below the cut-off grade (z ≤ zc ); (3) correct
rejection—if the block grade is below the cut-off grade (z ≤ zc ); and (4) false
rejection—if the block grade is above the cut-off grade (z > zc ).
There are two possible undiscounted dollar values for a block: the recovered
value and the potential value. The recovered value is the value that is recovered
given the block classification: (prz − cm − cp ) for a block called ore; and (−cm )
for a block called waste. The potential value is the value of the block if it was
correctly classified. The recovered and potential values are equal for correctly
classified blocks.
The indicator i (z, zc ) is defined as 1 for waste and 0 for ore:

1, z ≤ zc
i(z, zc ) = (2)
0, z > zc .

Estimation Methods

The simple polygonal method (Poly) consists of applying the cut-off grade on
polygonal (closest sample) block grades to classify them as ore or waste. In other
words, select as ore (ORE) if the closest sample value is larger than zc , otherwise
select as waste (WST).
The ordinary kriging approach (OK) consists of applying the cut-off grade

on kriged block grades to classify them as ore or waste. If zOK is the kriged block

grade, then select as ORE if zOK > zc , otherwise select as WST. Special attention
must be paid to the accuracy and precision of the kriged estimates. In particular,
there is always a conditional bias that must be reduced as much as possible by
using appropriate sample search criteria. The problem of conditional bias has been
documented in many publications (David, 1977, 1988; David and others, 1984;
Journel and Huijbregts, 1978; Krige, 1997).

Simulation-Based Methods

The average simulation method (AvgSim) consists of applying a cut-off


grade on the expected values of the block grades to classify them as ore or waste.
If Z is the random variable of the block grade and E[Z] its expected value, then
select as ORE if E[Z] > zc , otherwise select as WST. In theory, the block grade
expectation is conditional on the available information. In practice, it is obtained
by averaging several conditionally simulated values. Care must be taken to ensure
that the resulting averages are as close as possible to conditional expectations.
The minimum loss method (Isaaks, 1990; Srivastava, 1987) (MinLoss) con-
sists of computing the expected loss associated with each classification and picking
the classification that yields the minimum expected loss. For a given block, the
loss is defined by the potential value of the block minus the recovered value.
Downloaded form http://iranpaper.ir

454 Verly

The loss function (g(z)) of a block called ore and its expectation (LORE ) are:

(−cm ) − (prz − cm − cp ), z ≤ zc
g(z) =
(prz − cm − cp ) − (prz − cm − cp ), z > zc ,
LORE = E[g(Z)] = E[i(Z, zc ) × (−prZ + cp )]. (3)

The loss function (g(z)) of a block called waste and its expectation (LWST )
are:

(−cm ) − (−cm ), z ≤ zc
g(z) =
(prz − cm − cp ) − (−cm ), z > zc ,
LWST = E[g(Z)] = E[(1 − i(Z, zc )) × (prZ − cp )]. (4)

Select ORE if LORE < LWST , otherwise select WST. A geometrical interpre-
tation of the loss functions g(z) is given in Figure 1a. For a given block and category
(ore or waste), the expected loss is computed by averaging the simulated losses
corresponding to conditionally simulated block grades. The category that yields
the minimum average simulated loss is selected. The predicted loss of the block
is the average simulated loss of the selected category, that is, the minimum loss
value. The predicted grade of the block is the average simulation (AvgSim) value.
The maximum profit approach (MaxProf) is perhaps more intuitive than
minimum loss. It consists of computing the expected profit associated with each
classification and selecting the classification that yields the maximum expected
profit. For a given block, the profit is defined as the recovered value of the block.
The profit function (g(z)) of a block called ore and its expectation (PORE ) are:

g(z) = prz − cm − cp ,
PORE = E[g(Z)] = E[prZ − cm − cp ]. (5)

The profit function (g(z)) of a block called waste and its expectation (PWST )
are:

g(z) = −cm ,
PWST = E[g(Z)] = E[−cm ] = −cm . (6)

Select ORE if PORE > PWST , otherwise select WST. A geometrical inter-
pretation of the profit functions g(z) is given in Figure 1b. For a given block
and category (ore or waste), the expected profit is computed by averaging the
simulated profits corresponding to conditionally simulated grades. The category
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 455

Figure 1. Profit/loss functions corresponding to the minimum loss, maximum profit and
profit/loss approaches. In each case, there are four scenarios: correct/false acceptance for
blocks called ore, correct/false rejection for blocks called waste. The arrows in (c) ProfLoss
show the impact of the (ω1 , ω2 ) factors. MaxProf, ProfLoss-1|2|3 are particular cases of
ProfLoss ((ω1 , ω2 ) = (1, 0), (1, 0), (1, 1), and (2, 1), respectively).

that yields the maximum average simulated profit is selected. The predicted profit
of the block is the average simulated profit of the selected category, that is, the
maximum profit value. The predicted grade of the block is the average simulation
(AvgSim) value.
Downloaded form http://iranpaper.ir

456 Verly

The profit/loss approach (ProfLoss) as suggested by Glacken (1997) ignores


the mining cost cm and uses only one cut-off grade. This leaves room for in-
terpretation when formulating the profit in a more general case. The following
interpretation may not be unique but it handles several cut-off grades and is con-
sistent with Glacken’s examples:
• Profit = recovered value − (weight × penalty)
• Penalty = potential value − recovered value.
For a block called ore, the recovered value is (prz − cm − cp ). If the grade
is higher than the cut-off grade (correct acceptance):
• Penalty = 0
• Profit = prz − cm − cp .
If the grade is lower than the cut-off grade (false acceptance):
• Potential value = −cm
• Penalty = (−cm ) − (prz − cm − cp ) = −(prz − cp )
• Profit = (prz − cm − cp ) − ω1 (−(prz − cp )) = −cm + (1 + ω1 ) (prz − cp ),
where ω1 is the weight applied to the penalty. (1 + ω1 ) = ω1 is the formula-
tion used by Glacken (1997) who considers ω1 as a “coefficient quantifying
the impact of loss, and is related to the risk aversion profile of the company.”
ω1 is referred to as a risk aversion factor.
For a block called waste, the recovered value is (−cm ). If the grade is lower
than the cut-off grade (correct rejection):
• Penalty = 0
• Profit = − cm .
If the grade is higher than the cut-off grade (false rejection):
• Potential value = (prz − cm − cp )
• Penalty = (prz − cm − cp ) − (−cm ) = (prz − cp )
• Profit = −cm − ω2 (prz − cp ), where ω2 is the weight applied to the penalty.
Glacken (1997) refers to ω2 as a coefficient that quantifies “the willingness
of the company to accept wasting resource through false rejection of ore
blocks.” ω2 is referred to as a loss of opportunity cost factor.
The profit/loss function (g(z)) of a block called ore and its expectation (PLORE )
are:

−cm + ω1 (prz − cp ), z ≤ zc
g(z) =
prz − cm − cp , z > zc ,
PLORE = E[g(Z)] = E[i(Z, zc ) × ω1 (prZ − cp )]
+ E[(1 − i(Z, zc )) × (prZ − cp )] − cm . (7)
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 457

The profit/loss function (g(z)) of a block called waste and its expectation
(PLWST ) are:

−cm , z ≤ zc
g(z) =
−cm − ω2 (prz − cp ), z > zc ,
PLWST = E[g(Z)] = E[(1 − i(Z, zc )) × (−ω2 (prZ − cp ))] − cm . (8)

Select ORE if PLORE > PLWST , otherwise select WST. A geometrical inter-
pretation of the profit functions g(z) is given in Figure 1c.
Glacken (1997) presents an approach (referred to herein as ProfLoss-1 or
PL1 ) in which he sets the values for the risk aversion and loss of opportunity cost
factors as: ω1 = 1, ω2 = 0. In this case, the profit function (g(z)) of a block called
ore and its expectation (PL1ORE ) are:

g(z) = prz − cm − cp ,
PL1ORE = E[g(Z)] = E[prZ] − cm − cp . (9)

The profit/loss function (g(z)) of a block called waste and its expectation
(PL1WST ) are:

g(z) = −cm ,
PL1WST = E[g(Z)] = −cm . (10)

Select ORE if PL1ORE > PL1WST , otherwise select WST. A geometrical inter-
pretation of the profit functions g(z) is given in Figure 1d.
Deutsch and others (1999) suggest a simplification (referred to herein as
ProfLoss-2 or PL2 ) of Glacken’s profit/loss approach that amounts to setting
both risk aversion and loss of opportunity cost factors to 1 (ω1 = ω2 = 1). This
is equivalent to applying a penalty for under-estimation (false rejection), but no
penalty for over-estimation (false acceptance). Indeed, the false acceptance penalty
weight is ω1 such that:

(1 + ω1 ) = ω1 ⇔ ω1 = 0, if ω1 = 1. (11)

The argument to penalize under-estimation is that “it does cost money to


mistakenly put high grade ore on the waste dump” (Deutsch and others, 1999).
The profit function (g(z)) of a block called ore and its expectation (PL2ORE ) are:

g(z) = prz − cm − cp ,
PL2ORE = E[g(Z)] = E[prZ] − cm − cp . (12)
Downloaded form http://iranpaper.ir

458 Verly

The profit function (g(z)) of a block called waste and its expectation (PL2WST )
are:

−cm , z ≤ zc
g(z) =
−cm − (prz − cp ), z > zc ,

PL2WST = E[g(Z)] = E[(1 − i(Z, zc )) × (−prZ + cp )] − cm . (13)

Select ORE if PL2ORE > PL2WST , otherwise select WST. A geometrical inter-
pretation of the profit functions g(z) is given in Figure 1e.
Another simplification (referred to herein as ProfLoss-3 or PL3 ) consists
of penalizing both under- and over-estimation, which could make sense since
money is lost in both cases. This simplification amounts to setting the weights
as: ω1 = ω2 = 1, or ω1 = 2 and ω2 = 1. The profit/loss function (g(z)) of a block
called ore and its expectation (PL3ORE ) are:

−cm + 2(prz − cp ), z ≤ zc
g(z) =
prz − cm − cp , z > zc ,

PL3ORE = E[g(Z)] = E[i(Z, zc ) × 2(prZ − cp )]


+ E[(1 − i(Z, zc )) × (prZ − cp )] − cm . (14)

The profit/loss function (g(z)) of a block called waste and its expectation
(PL3WST ) are:

−cm , z ≤ zc
g(z) =
−cm − (prz − cp ), z > zc ,

PL3WST = E[g(Z)] = E[(1 − i(Z, zc )) × (−prZ + cp )] − cm . (15)

Select ORE if PL3ORE > PL3WST , otherwise select WST. A geometrical inter-
pretation of the profit functions g(z) is given in Figure 1f.

COMPARISON OF THE METHODS

The simple polygonal approach bases its classification on polygonal esti-


mates. These are often the worst possible estimates, so this approach is expected
to perform poorly.
The ordinary kriging and average simulation methods are similar in the
sense that both apply the same cut-off grade on average grade maps. The average
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 459

grades are different because the assumptions/algorithms are different. In theory,


average simulation is superior to ordinary kriging because it is a conditional
expectation whereas ordinary kriging is conditionally biased. In practice, however,
ordinary kriging is a robust estimator, whereas the assumptions needed for average
simulation to get a conditional expectation (e.g., multi-normality) have sometimes
to be stretched unreasonably. Not much difference would be expected between the
two methods in a grade control situation because numerous data are available. The
case studies presented in this paper, however, show that this may not be always
the case in practice.
The maximum profit approach is optimum in the sense that it selects the
classification that corresponds to the maximum profit. In this case, the profit
is a linear function of the grade but it does not need to be so. The formula-
tion is perfectly general and can include any bonuses and/or penalties that may
apply.
The average simulation and maximum profit methods are equivalent in the
case where profits are linear functions of the grade (Fig. 1b). Maximum profit calls
a block ore if (Eqs. (5) and (6)):

PORE > PWST


⇔ E[prZ − cm − cp ] > −cm ⇔ prE[Z] > cp ⇔ E[Z] > cp /pr
⇔ E[Z] > zc , i.e., the criteria for average simulation. (16)

The maximum profit and minimum loss methods are equivalent. Maximum
profit calls a block ore if:

PORE > PWST


⇔ E[prZ − cm − cp ] > −cm
⇔ E[i(Z, zc ) × (prZ − cp )] + E[(1 − i(Z, zc )) × (prZ − cp )] > 0
⇔ E[(1 − i(Z, zc )) × (prZ − cp )] > 0 − E[i(Z, zc ) × (prZ − cp )]
⇔ LWST > 0 − (−LORE ), cf. Eqs. (3) and (4)
⇔ LORE < LWST , i.e., the criteria for minimum loss. (17)

At first, the profit/loss approach appears attractive because of its generality.


However, a proper choice of the risk-aversion loss (ω1 ) and opportunity cost loss
(ω2 ) factors is not easy. The first “obvious” choice is ω1 = 1, ω2 = 1, as for
ProfLoss-2. Alternative choices are ω1 = 1, ω2 = 0, as for ProfLoss-1, or ω1 = 2,
ω2 = 1, as for ProfLoss-3.
Downloaded form http://iranpaper.ir

460 Verly

For the ProfLoss-1 case, the profit/loss and maximum profit methods are
equivalent. ProfLoss-1 calls a block ore if (Eqs. (9) and (10)):

PL1ORE > PL1WST


⇔ E[prZ] − cm − cp > −cm
⇔ prE[Z] > cp ⇔ E[Z] > cp /pr (18)
⇔ E[Z] > zc ⇔ PORE > PWST ,
i.e., the criteria for maximum profit (Eq. (16)).

This conclusion can also be reached by comparing the geometrical interpreta-


tions of the methods. In Figure 1c (ProfLoss), setting ω1 = 1, ω2 = 0 is equivalent
to rotating the profit segments corresponding to false acceptance and false rejec-
tion for a result as per Figure 1d (ProfLoss-1), which is the same as per Figure 1b
(MaxProf).
In looking at the ProfLoss-2 case, Figure 1b (MaxProf) and e (ProfLoss-2)
show that the only difference between the two sets of functions occurs for blocks
called waste that are in fact ore (false rejection). In this case, the profit/loss
approach is including a penalty and, as a result, will identify less waste and more
ore than the maximum profit approach. The maximum profit approach already
penalizes misclassification (false acceptance profit is less than correct rejection
profit, and false rejection profit is less than correct acceptance; Fig. 1b). The
penalty imposed by ProfLoss-2 for false rejection (Fig. 1e) can be viewed as an
additional penalty that is hard to justify. Indeed, the mining operation is not paying
a fine for misclassifying ore as waste and, if it was, this should be built into the
maximum profit formulation. The ProfLoss-2 approach is therefore expected to
not perform as well as maximum profit in a usual grade control situation.
ProfLoss-3 is the case that applies a penalty (unweighted) to both under-
and over-estimation. In this case, the profit/loss and maximum profit methods are
equivalent. ProfLoss-3 calls a block ore if (Eqs. (14) and (15)):

PL3ORE > PL3WST


⇔ E[i(Z, zc ) × 2(prZ − cp )] + E[(1 − i(Z, zc )) × (prZ − cp )] − cm
> E[(1 − i(Z, zc )) × (−(prZ − cp ))] − cm
⇔ 2E[i(Z, zc ) × (prZ − cp )] + 2E[(1 − i(Z, zc )) × (prZ − cp )] > 0 (19)
⇔ E[prZ − cp ] > 0
⇔ PORE > PWST , as already established for ProfLoss-1 (Eq. (18)).
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 461

In summary, the comparison of grade control methods has assumed that the
simple polygon method is the worst performer in the simple case of one cut-
off grade and the recovered profit being a linear function of the grade. Average
simulation is better than ordinary kriging in theory, but probably equivalent in
practice. Average simulation, minimum loss and maximum profit approaches are
equivalent. The maximum profit approach is optimum, which implies that average
simulation is optimum too. Two of the profit/loss approaches (ProfLoss-1 and
ProfLoss-3) are equivalent to maximum profit. The other profit/loss approach
(ProfLoss-2) is not expected to perform as well as maximum profit, as it identifies
more waste as ore and vice versa.
In the case of several cut-off grades, similar results are noted. Differences are
noted if the recovered profit is not a linear function of the grade. In this case, the
maximum profit approach is the only one that is optimum.

WALKER LAKE CASE STUDIES

First Case Study

To further illustrate the similarities and differences between the various meth-
ods, a case study on the Walker Lake dataset, as suggested by Isaaks (1990) and
Dimitrakopoulos (2001), has been completed. The dataset covers a lateral extent
of 260 m × 300 m. The “truth” consists of 780 true 10 m × 10 m block grades
(Fig. 2c). A total of 195 blast hole sample values on a 20 m × 20 m grid are used
for simulation and kriging (Fig. 2a). There are four ore categories: Waste, ROM
(run-of-mine ore), Agg (aggregate ore), and Mill (mill ore). The economic param-
eters are given in Table 1. The profit functions per category are given in Table 2.
Ore-waste selection is based on all the grade control methods discussed above
(i.e., Poly, OK, AvgSim, MinLoss, MaxProf, and ProfLoss-1|2|3). There are no
smoothed digging lines; each 10 m × 10 m block can be selected as Waste, ROM,
Agg or Mill ore.
Figure 2 shows location and ore-category maps together with some statistics.
Recovered tons, grade and value per category are given in Table 3. A profit room is
defined as the maximum profit that can be achieved by improving on the simplest,
but also worst, polygonal grade control method. In other words, it is the true
recovered value minus the Poly recovered value (3180.1 − 1712.9 = 1467.2, see
Table 3). Table 4 lists incremental values between the grade control methods and
the corresponding percentages of the profit room.
Because it has four ore categories, this case study is more complex than the
two-category situation considered earlier. However, the profits per category are
linear functions of the grade (Table 2), which means that the conclusions reached
in the previous section are unchanged. Tables 3 and 4 show that Poly produces
by far the worst value (1712.9), which is expected since the polygonal method
Downloaded form http://iranpaper.ir

462 Verly

Figure 2. Walker Lake first case study. Location maps, statistics, and true and estimated ore
category maps.

produces bad estimates, hence bad classifications. There is a 35% increase in value
from Poly to OK (1712.9–2305.7), which corresponds to 40.4% of the profit room.
There is a further 6.4% increase in value from OK to AvgSim (2305.7–2454.2),
which corresponds to 10.1% of the profit room (7.2 + 2.9%; see Table 4). Such
an increase would be very significant for an operation; but the result is surprising
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 463

Table 1. Walker Lake Economic Parameters

Waste ROM Aggregate ore Mill ore

Mining cost 1.50 1.50 1.50 1.50


Processing cost 0 1.44 3.64 8.64
Min + proc cost 1.50 2.94 5.14 10.14
Recovery 0% 45% 70% 95%
Grade range (opt) 0–0.008 0.008–0.022 0.022–0.05 0.05+

Note. Gold price is 400 USD per ounce. Costs are in USD per (short) ton. Grade
ranges are in ounce per ton (opt).

since there is no compelling reason for AvgSim to outperform OK. In fact, Table 3
and Figure 2 show that there are some very significant differences between the OK
and AvgSim ore categories. This is probably related to the difficulty in estimating
or simulating Walker Lake (a mountain range and a lake), and also to a lack of
information. Perhaps, also, AvgSim outperforms kriging for its ability to better
account for skewed distributions. It is worth noting that sample values are available
on approximately a 6 m × 6 m grid in the two case studies of operating mines
discussed later in this paper, rather than the 20 m × 20 m grid for Walker Lake.
In these actual cases, much less difference is noted between the OK and AvgSim
ore-category maps.
As predicted by theory, the same results are obtained for AvgSim, MinLoss,
MaxProf, ProfLoss-1, and ProfLoss-3. There is a 1.7% decrease in value from
AvgSim to ProfLoss-2 (2454.2–2411.6; see Table 3), which was expected since
ProfLoss-2 tends to perform worse than MaxProf (and AvgSim) because of the
additional penalty for false rejection. The additional penalty causes ProfLoss-2
to identify less waste and more ore than MaxProf and this can be seen in
Table 3. ProfLoss-2 identifies 29% fewer Waste tons (10.4% compared with
14.6%) and 51% more Mill ore tons (12.8% compared with 8.5%) than
MaxProf.

Second Case Study

In the first study, ProfLoss-2 performs slightly worse than AvgSim. To il-
lustrate the fact that ProfLoss-2 generally under-performs AvgSim, the recovered
values in both methods are now computed using one realization at a time for

Table 2. Profit Function (per Ton) for Each Ore Category

Waste ROM Aggregate ore Mill ore

Profit function −1.50 180 × z − 2.94 280 × z − 5.14 380 × z − 10.14


Downloaded form http://iranpaper.ir

464 Verly

Table 3. Recovered Tons (%), Grade (Opt), and Value (USD) for the Grade Control Methods

Waste ROM ore Aggregate ore Mill ore Combined

Truth %Tons 34.1 26.0 23.5 16.4 100


Grade 0.0026 0.0147 0.0336 0.0854 0.0266
Value −399.0 −60.4 783.5 2856.1 3180.1
Poly %Tons 57.4 16.9 9.2 16.4 100
Grade 0.0141 0.0309 0.0324 0.0628 0.0266
Value −672.0 346.6 282.5 1755.8 1712.9
OK %Tons 22.4 40.8 27.1 9.7 100
Grade 0.0110 0.0184 0.0292 0.0892 0.0266
Value −262.5 121.1 642.8 1804.2 2305.7
AvgSim %Tons 14.6 45.4 31.5 8.5 100
Grade 0.0075 0.0159 0.0337 0.0904 0.0266
Value −171.0 −27.9 1055.6 1597.5 2454.2
MinLoss As per AvgSim
MaxProf As per AvgSim
ProfLoss-1 As per AvgSim
ProfLoss-2 %Tons 10.4 39.5 37.3 12.8 100.0
Grade 0.0074 0.0140 0.0289 0.0742 0.0266
Value −121.5 −130.8 856.8 1807.1 2411.6
ProfLoss-3 As per AvgSim

the truth (instead of the Walker Lake true 10 m × 10 m block ore grades). The
resulting 100 pairs of “true” recovered profits are shown in Figure 3. The figure
shows that AvgSim outperforms ProfLoss-2 most of the time (83 times out of
100). On average, the percent change in value from AvgSim to ProfLoss-2 is a
2.1% decrease (1857.4–1819.2).
ProfLoss-2 tends to outperform AvgSim for the higher recovered profits
(17 times out of 100), which can be explained as follows. Higher profits correspond
to higher-grade realizations. These higher-grade realizations are “optimistic” with
respect to the 195 blast hole sample values. In other words, they contain more ore
and less waste than indicated by the sample values. It makes sense then to have

Table 4. Incremental Values of the Various Grade Control Methods and Corresponding Percentages
of the Profit Rooma

Poly → OK → ProfLoss-2 → AvgSim → Poly →


OK ProfLoss-2 AvgSim Truth Truth

Incremental value 592.8 105.9 42.6 725.9 1467.2a


Incremental percentb 40.4% 7.2% 2.9% 49.5% 100.0%

a The profit room is the true value minus the poly grade control recovered value (3180.1 − 1712.9 =
1467.2, see Table 3).
b The incremental percent is the percentage of the 1467.2 profit room.
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 465

Figure 3. Walker Lake second case study. AvgSim versus ProfLoss-2 recovered profits for 100
10 m × 10 m block grade realizations (100 simulated “truths”).

ProfLoss-2 more successful than AvgSim for these realizations since ProfLoss-2
identifies less waste and more ore than AvgSim.
The average of the 100 AvgSim recovered profits is 1857.4, that is, 24%
less than the actual recovered value (2454.2, for AvgSim in Table 3). Part of the
difference is due to the original average of the 195 blast hole samples (0.0236 opt),
which is 11% lower than the true average grade (0.0266 opt; see Table 3). Part of
the difference could also be a coincidence (10 realized AvgSim profits are higher
than 2454.2) or could simply be an indication that the simulation does a bad job at
reproducing the Walker Lake “mineralization”. This, however, has no bearing upon
the relative performance of AvgSim compared with ProfLoss-2. Note that if such
lack of reconciliation were observed in reality, immediate action would be needed.

ACTUAL CASE STUDIES AT MINE SITES

A few years ago, Placer Dome Inc. decided to investigate various means
for improving grade control. At first, most of the effort was put into using the
maximum profit approach using simulated grade values. It was soon realized,
however, that this approach was equivalent to simply applying the proper cut-off
Downloaded form http://iranpaper.ir

466 Verly

Figure 4. Cortez gold recovery curve for K1 = 0.8 and K2 = 0.001. The K1 and K2 values have
been modified for confidentiality reasons.

grades on the average of the simulated values when the profit is a linear function
of the grade. For this reason, the project evolved into looking at whatever is
required to improve grade control at operating mines. Two tests were conducted
at two operating mines: Cortez and Porgera. The main results of the two tests are
reported below.

Study at Cortez Mine, Nevada, USA

The Cortez Pipeline deposit is a disseminated Carlin-type gold deposit located


in Nevada, USA. Benches are 20 ft high. Blast hole samples are approximately
on a 20 ft grid. The ore-waste categories of interest are Waste, Leach, and Oxide
(mill) ore. The profit per category is a linear function of the grade even though the
recovery per category is not always constant. Indeed, the mill ore gold recovery r
is modeled as a nonlinear function of the block grade z (Fig. 4):
K2
r = K1 − , (20)
z
where K1 and K2 are two constants. The profit per ton, however, is a linear function
of the grade:
P = rpz − cp = K1 pz − K2 p − cp , (21)

where p is the gold price and cp the processing cost per ton.
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 467

It is worth noting that with such a model, the mill gold recovery becomes
negative when the grade gets close to 0. The grade, however, is then so low with
respect to the leach/mill cut-off grades that this has no impact on the leach/mill
classification.
Before the test, kriging was not used at Cortez for grade control. The reason
cited was that kriging was “spreading the grades.” Digging lines were simply drawn
between color-coded blast hole values without too much attention to geological
trends.
A test was performed in 2000 on four benches to investigate the differences
between various grade control approaches. Two polygonal grade control methods
were investigated: Poly1, which stands for the old grade control practice at the
mine; and Poly2, which pays more attention to geological trends when design-
ing digging lines. Other approaches examined were AvgSim, with digging lines
based on average simulated grade maps (sequential Gaussian simulation) and at-
tention paid to geological trends; OK, the same but with digging lines based on
kriged grade maps (ordinary kriging); and MaxProf, the maximum profit approach
described in this paper.
For the AvgSim, OK, and MaxProf methods, the four benches were divided
into 47 smaller “zones.” A typical zone is approximately 300 ft × 300 ft and con-
tains more than 300 blast hole values per bench (Fig. 5b). Some zones were further
divided into two domains. The objective of the smaller zones/domains is to get
better local statistics, in particular better variograms, and to account as much as
possible for local geology. Correlograms at Cortez are easy to fit with very low
nugget values and long ranges (Fig. 5e). The anisotropy directions, however, can
significantly change locally (Fig. 5c). Gold grades were simulated/estimated on a
2 ft × 2 ft grid using blast hole values from the current bench and the bench above.
Sixteen blast hole values were used per kriged grid node. Sixteen conditioning
values (blast holes and/or previously simulated values) were used per simulated
grid node.
For comparison purposes, the average simulated grade map (AvgSim) was
selected as the “truth” (Fig. 5g). This map is the same as the one used to test the
AvgSim approach. It therefore gives an advantage to all approaches based on the
same simulation algorithm (AvgSim and MaxProf) over those that are not (Poly1,
Poly2, and OK). The generated “truth” is far from perfect. Still, it was considered
the best that could be done given the available blast hole values and geological
information (a better “truth” could have been obtained by also using the blast hole
samples from the bench below along with drill hole samples; this was done for the
second test at Porgera). Maps of the “truth” showed many global and local trends
that were considered very realistic by Cortez geological staff. Several cursory
comparisons were also made with four different gold grade realizations used for
the truth (Fig. 5f). The results were obviously different, but not very much, and all
the findings about the merits of the approaches, including their ranking in term of
Downloaded form http://iranpaper.ir

468 Verly

Figure 5. Cortez case study Zone 9. The zone is 400 ft × 300 ft (a). Grid lines are 200 ft apart (b).
Variograms are continuous (e) but with variable directions as shown by variogram maps in (c). One
variable is considered: Au (d, e). The estimated ore category maps are shown in (g). The truth is
identical to the estimated AvgSim (and MaxProf) map. The different ore blockouts are shown in (h).
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 469

dollar value remained the same. Ranking details and findings are discussed further
below.
The profit per ore category at Cortez is a linear function of the grade
(Eq. (21)). The MaxProf and AvgSim approaches provide therefore identical
results (Fig. 5g).
Not surprisingly, the Poly1 and Poly2 results showed that improving ore block
limits by accounting for geological trends has a significant dollar value. The other
significant improvement is obtained by using better grade maps (e.g. AvgSim or
OK) and improved digging lines (more attention to geological trends). The dollar
value of AvgSim and improved digging lines over the older Poly1 method is up
to 60% of the profit room (as defined in the Walker Lake first case study) for the
type of ore complexity encountered during the test. This value is equivalent to
approximately 1 million USD per year or about 1 USD per ounce of gold.
The AvgSim dollar value is slightly better than the OK dollar value (about
5% of the profit room). However, this improvement is not considered significant
since the reference for the comparison—the “truth”—is itself simulated.
Significant mineralization trend changes can be observed in the same geolog-
ical unit but cannot always be predicted from one bench to the next. Computing
variograms locally per blast zone, and not globally at the level of a few benches,
was considered useful to improve the estimated maps (AvgSim or OK).

Study at Porgera Mine, Papua New Guinea

The Porgera deposit is an epithermal gold deposit located in the highlands of


Papua New Guinea. Base metal carbonate and quartz roscoelite gold mineralization
occur as two distinct phases within a complex structural framework of cross-cutting
faults, splays and link structures. Benches are 10 m high. Blast hole samples are
approximately on a 6 m grid (Fig. 6b). The ore categories are Waste, Blue Ore
(stockpile), and Red Ore (mill). Metallurgy is complex, and sulfur has a negative
impact on gold recovery and costs. Profit is not a linear function of the gold grade;
its formulation is complex and not reported in this paper.
Kriging was performed for grade control at Porgera before the test. However,
the 5 m × 5 m kriged block grades were judged too coarse by the geological staff.
For this reason, ore blocks were manually drawn between color-coded blast hole
values. Kriging results were used to compute average grades within the manually
designed ore blocks. The kriging parameters (variograms, trimming values, etc.)
had not been revised for a few years and the same variogram model was used
whatever the location in the deposit.
A test was performed in 2001/2002 on two benches to investigate the dif-
ferences between various grade control approaches. The approaches investigated
were Poly, which stands for the old grade control practice; AvgSim, with digging
lines based on average simulated grade maps (Au and S, sequential Gaussian co-
simulation); OK, the same as AvgSim, but using a grid of kriged values instead
Downloaded form http://iranpaper.ir

470 Verly

Figure 6. Porgera case study Zone 1. The zone is 70 m × 130 m (grid lines are 50 m apart)
and consists of two geology domains (a, b). Two variables are considered: Au and S (c).
The estimated ore category maps are shown in (f). More data have been used to generate
the “truth” (g), which is different from the estimated AvgSim map (f). The ore blockouts are
shown in (h). Note that the original Poly ore blockout was particularly poor for Zone1.
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 471

(ordinary kriging for each of Au and S); and MaxProf, the maximum profit ap-
proach described earlier, using co-simulated Au and S values.
Sub-vertical high-grade veins at Porgera are easily missed by vertical blast
holes. For this reason, the impact of using drill hole samples in addition to blast
hole samples was also investigated.
Geological staff at Porgera believes that the structural complexity and cross-
cutting relations of many mineralized structures require a “local” approach for
better estimation or simulation. For the AvgSim, OK, and MaxProf methods, the
two benches were divided into 21 smaller “zones.” A typical zone is approximately
100 m × 100 m and contains about 300 samples (blast holes and drill holes) per
bench. Each of the zones was further subdivided into two and sometimes three
geology domains (Fig. 6b).
The drill hole and blast hole samples were pooled into a single dataset when
the two sample types were used. Au and S grades were estimated or co-simulated
on a dense 0.625 m × 0.625 m grid using blast hole values from the current bench
and the bench above. Sixteen blast hole values were used per kriged grid node.
Sixteen conditioning values (blast holes and/or previously simulated values) were
used per simulated grid node. Correlograms were used for kriging. Models of linear
corregionalization were used for co-simulation. The normal score variograms for
simulation (in fact correlograms) were often easier to fit than the sometimes noisy
correlograms used for kriging (Fig. 6d).
As for Cortez, average co-simulated grade Au and S maps were used to
produce “true” ore category maps (Fig. 6g). These maps have been produced
using blast hole and drill hole samples from the current bench and the benches
above and below. The “true” maps are different from the maps used to test the
AvgSim approach, which were produced without information from the bench
below (Fig. 6f). Still, the “truth” consists of average co-simulated maps, which
gives an advantage to all approaches based on a similar co-simulation algorithm
(AvgSim and MaxProf) over those that are not (Poly and OK). This was considered
when interpreting the test findings that are discussed below.
The profit for a given category at Porgera is not a linear function of the Au
and S grades. The MaxProf and AvgSim approaches therefore provide different
ore category maps, with MaxProf theoretically being better than AvgSim. The
test, however, showed that the differences between the AvgSim and MaxProf are
limited to a few scattered cells on the ore category maps, resulting in insignificant
dollar value for perfect selection and no dollar value after ore blocking (Fig. 6f).
Ore block boundaries at Porgera are complex (Fig. 6h). The test showed
that a significant dollar value is achieved by locating the boundaries more accu-
rately, which can be accomplished by using OK or AvgSim instead of polygonal
estimates. The observed dollar value increase in switching from Poly to OK or
AvgSim is in the order of several million USD per year for the material considered
for the test. Adding drill hole information to blast hole information also results
in additional dollar value. On one of the two benches, there is no advantage in
Downloaded form http://iranpaper.ir

472 Verly

switching from OK to AvgSim. On the other bench, however, there is a significant


difference (about 30% of the profit room). The geological staff reviewed the OK
and AvgSim ore category maps and preferred the AvgSim maps, as they seem to
better represent the sharp structural contacts observed.
Among other findings, the test showed that ore blocking on coarse grid
maps (5 m × 5 m, Fig. 5h), as was done and rejected before the test, is not much
worse than on finer grid maps (0.625 m × 0.625 m). However, the finer grid maps
look more realistic, with sharp, well-defined boundaries, and therefore have more
chance of being properly used by the staff. Finally, before the test, polygonal
ore blocking was very inconsistent. Using OK or AvgSim maps results in more
consistency and less misclassification.

DISCUSSION AND CONCLUSIONS

Among the simulation-based grade control approaches discussed in this pa-


per, maximum profit is considered to be the best. Indeed, it is somewhat simpler
than minimum loss, to which it is equivalent, and it is definitely simpler than the
profit/loss approaches. It accounts for loss of revenue due to under- and over-
estimation without over-penalizing one or the other as profit/loss can do (for
example, the ProfLoss-2 approach). Finally, maximum profit is theoretically bet-
ter than average simulation if the recovered profit is a nonlinear function of the
grade for a given category.
In a typical grade control situation the recovered profit per category can often
be expressed as a linear function of the grade, even if the metal recovery is not con-
stant (cf. Cortez Pipeline). In such a case, maximum profit and average simulation
result in the same classification. If the recovered profit is not a linear function of the
grade, the two approaches result in different classifications and maximum profit
is optimal. In practice, however, it should be verified that the differences are not
insignificant or spurious. Insignificant differences were noted at Porgera. Spurious
differences would be noted if the nonlinearity were due to an artifact of metal-
recovery fitting, i.e. another reasonable fit would result in a linear profit function
(as was the case for Cortez South Pipeline, which was not used for the test).
If maximum profit and average simulation are equivalent in results, the next
choice is between average simulation and ordinary kriging. For Walker Lake,
average simulation is 10% better in terms of profit room than ordinary kriging.
For Cortez, average simulation is 5% better, which was not considered significant
because the test was biased toward simulation. For Porgera, average simulation
is equivalent to ordinary kriging for one bench, but is 30% better than ordinary
kriging for the other bench. This was considered quite significant even if the test
was biased toward simulation. Porgera staff indicated their preference for average
simulation, as it seems to better represent the sharp structural contacts observed.
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 473

Average simulation is theoretically better than ordinary kriging if the sim-


ulation assumptions (multi-normal assumption after transformation for Gaussian
simulation) are valid. In particular, average simulation is better than ordinary
kriging at handling skewed distributions. In practice, however, ordinary kriging is
more robust than average simulation because it only requires local stationarity.
More work is needed on the choice between ordinary kriging and average
simulation, and between the various simulation methods. It is important to use
examples carefully when doing this work. When the “truth” is simulated as it was
for the Cortez and Porgera examples, an unfair advantage is given to methods that
use similar algorithms. If the “truth” is a realization, the relative performance of
simulation-based methods may depend on the realization itself (as shown in the
second Walker Lake study).
One point that cannot be contested is the significant increase in dollar value
that can be achieved by switching from polygonal estimates to some average grade
maps (ordinary kriging or average simulation). This is not news to geostatisticians.
However, it is surprising how much the polygonal method is still in use for grade
control in many operations. Additional dollar value can be obtained by using
all relevant information (e.g., drill hole information in the Porgera study) and
designing smoothed consistent digging lines that account for mineralization trend
(as in Cortez and Porgera). Once these additive advantages are achieved, it is not
clear why average simulation should be better than ordinary kriging in terms of
dollar value if profits are linear functions of the grades.
Simulation has several other advantages over kriging that have not been dis-
cussed. This is because their dollar values are hard to quantify in a context of
daily grade control that must be automated as much as possible. For example,
simulation offers more opportunities to incorporate geological input, soft infor-
mation, and spatial trends. Simulated maps are not smoothed like kriged maps
and are particularly suited for a specific geological input, for example, the visual
judgment of the geologists. The problem of change of support is easily solved
through simulation. Simulation is much better than kriging for quantifying risk
with, for example, probability maps that are readily accessible. Simulation can
also be used to identify the fixed tonnage that has a better chance to exceed a
cut-off grade. Finally simulation is a great tutorial tool because it is so visual.
If geostatistics is used (simulation or kriging), it must be applied correctly
otherwise mining operations quickly revert to polygonal estimates. In the case
of Cortez and Porgera, “good” geostatistics was achieved by having a set of
geostastistical parameters for local zones that correspond to a couple of days of
production.
Michel David was renowned for his practical touch. This paper has shown
that most of the significant improvements in grade control are achieved through
the simple rules that David always preached: good samples, good geology and, of
course, good practical geostatistics.
Downloaded form http://iranpaper.ir

474 Verly

ACKNOWLEDGMENTS

The author wishes to thank Placer Dome Inc., Cortez Gold Mines, and Porgera
Joint Venture for permission to publish this paper. He is also grateful to the many
people who made this project possible, notably Britt Buhl of Cortez, Elizabeth
Haren of Granny Smith, Tony Burgess and Jim Ged of Porgera, Graham Barber,
John Brunette, and Dick Lewis of the Vancouver and Sydney offices. Finally, he
wishes to thank Mohan Srivastava and an anonymous reviewer for their useful
comments and suggestions.

REFERENCES

Collet, D., and Corley, D., 2000, Transparent grade control: Improvements from the application of
sequential indicator simulation and optimisation at Ravenswood gold mines, NQ, in CD-ROM,
International Symposium on Geostatistical Simulation in Mining, 1999, Perth, WA, Australia.
David, M., 1977, Geostatistical ore reserve estimation: Elsevier, Amsterdam, 364 p.
David, M., 1988, Advanced applied geostatistical ore reserve estimation: Elsevier, Amsterdam,
217 p.
David, M., Marcotte, D., and Soulié, M., 1984, Conditional bias in kriging and a suggested correction,
in Verly, G., David, M., Journel, A. G., and Maréchal, A., eds., Geostatistics for natural resource
chracterisation: D. Reidel Publishing Co, Dordrecht, The Netherlands, Vol. 1, p. 217–230.
Deutsch, C. V., Magri, E., and Norrena, K., 1999, Optimal grade control using geostatistics and
economics: Methodology and examples. SME Annual Meeting, Denver, March 1999.
Dimitrakopoulos, R., 2001, Geostatistical simulations for the mining industry. Course Notes, WH
Bryan Mining Geology Research Centre, Brisbane, 309 p.
Goovaerts, P., 1997, Geostatistics for natural resources evaluation: Oxford University Press, New York,
483 p.
Glacken, I. M., 1997, Change of support and use of economic parameters for block selection, in Baafi,
E. Y., and Schofield, N. A., eds., Geostatistics Wollongong ’96: Kluwer Academic Publishers,
Dordrecht, The Netherlands, Vol. 2, p. 811–821.
Godoy, M. C., Dimitrakopoulos, R., and Costa, J. F., 2001, Economic functions and conditional
simulation applied to grade control. Mineral resource and ore reserve estimation—The AusIMM
guide to good practice. AusIMM Monograph 23, p. 591–599.
Gomez-Hernandez, J., and Journel, A. G., 1993, Joint sequential simulation of multiGaussian fields, in
Soares, A., ed., Geostatistics Troia ’92: Kluwer Academic Publishers, Dordrecht, The Netherlands,
Vol. 1, p. 85–94.
Isaaks, E. H., 1990, The application of Monte Carlo methods to the analysis of spatially correlated
data: Unpublished Ph.D. Thesis, Stanford University, Stanford, California, 213 p.
Isaaks, E. H., and Srivastava, R. M., 1989, An introduction to applied geostatistics: Oxford University
Press, New York, 561 p.
Journel, A. G., and Huijbregts, C. J., 1978, Mining geostatistics: Academic Press, London, 600 p.
Krige, D. G., 1997, A practical analysis of the effects of spatial structure and of data available and
accessed, on conditional biases in ordinary kriging, in Baafi, E. Y., and Schofield, N. A., eds.,
Geostatistics Wollongong ’96: Kluwer Academic Publishers, Dordrecht, The Netherlands, Vol. 2,
p. 799–810.
Shaw, W. J., and Khosrowshari, S., 1997, Grade control sampling and ore blocking: Optimisation
based on conditional simulation: 3rd International Mining Geology Conference, AusIMM-AIG,
Series 6.97, p. 131–134.
Downloaded form http://iranpaper.ir

Grade Control Classification of Ore and Waste 475

Srivastava, R. M., 1987, Minimum variance or maximum profitability?: CIM Bull., v. 80, no. 901,
p. 63–68.
Schofield, N., and Rolley, P., 1997, Optimisation of ore selection in mining: Method and case studies:
3rd International Mining Geology Conference, AusIMM-AIG, Series 6.97, p. 93–97.
Verly, G., 1993, Sequential Gaussian co-simulation: A simulation method integrating several types of
information, in Soares, A., ed., Geostatistics Troia ’92: Kluwer Academic Publishers, Dordrecht,
The Netherlands, Vol. 1, p. 543–554.

You might also like