Annual Report 2022

Paper 2022-04

A review of grade control methods in open cast mining1

Noble E. Potakey ([email protected])
Julian M. Ortiz ([email protected])

Abstract
Ore grade control is an important part of a mine’s short-term planning to define which
material is considered as ore or waste. The effectiveness of grade control is subject to
a compendium of factors such as sampling errors, conditional bias introduced by grade
estimators, accurate definition of dig limits, and blast-induced rock movements. In this
scenario, misclassification of the mined material which is mainly ascribable to the lack
of absolute knowledge about real grade distribution, is our major concern. This paper
reviews the state-of-the-art grade control practices used for classifying ore and waste
in open pit mining. Common approaches include the classical and distance weighting
estimation techniques, geostatistical methods such kriging, and simulation-based
methods. Theoretical review shows that conditional simulation is a better classifier of
ore and waste than estimation methods due to its ability to account for grade
uncertainties and the different optimization algorithms it provides to access economic
consequences of ore/waste decisions. Subsequently, a novel machine learning
approach using Elliptical Radial Basis Function Network (ERBFN) and Support Vector
Regression (SVR) was discussed. Results from a case study show that SVR achieved an
8%, 1.12% and 1.16% reduction in misclassified material relative to inverse distance,
ordinary kriging, and simulation respectively. The ERBFN model also obtained a
decrease in misclassified material of 12%, 5.4% and 5.7% compared to inverse distance,
ordinary kriging, and simulation-based approaches, respectively.

1. Introduction
Considering the extensive nature of mining operations which involves making decisions regarding large
volumes of materials taking place over limited period of times, blending of waste with ore and ore with
waste is inevitable. Nevertheless, geologists and mine engineers must ensure that this situation is reduced
to the barest minimum during excavation, and that ore and waste are differentiated. The decision of which
material is ore or waste is very crucial for the mine’s profitability since it is the last opportunity for the
mining company to achieve its estimated revenue, and errors at this stage are very costly and difficult to
reverse due to its proximity to the production stage (Rossi & Deutsch, 2014). To help make better
decisions in classifying ore and waste, and to select the destination of each parcel of material mined,
mines perform grade control.

Ore grade control is a compendium of procedures and practices usually involving blast hole sampling,
grade estimation, ore/waste classification, blast-induced rock movement measurements, defining dig
limits among others. It is done to identify which material is ore or waste and to ensure that the mill is fed

1
Cite as: Potakey N. E., Ortiz J. M. (2022) A review of grade control methods in open cast mining, Predictive
Geometallurgy and Geostatistics Lab, Queen’s University, Annual Report 2022, paper 2022-04, 59-70.

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 59

with the right grade of material. In this scenario, material classification based on the grade assignment is
discussed. A volume of material is classified as ore or waste based on the grade assigned to it from grade
prediction methods and by comparison to a cut-off grade (Abzalov, 2016). When the grade assigned to a
material falls below a given cut-off grade, it is considered as waste and sent to the waste dump while
material with grade above the cut-off grade is sent to the mill. However, when the assigned grades are
not consistent with the actual grade distribution, it results in a misclassification, and materials are sent to
the wrong destination. The figure below shows the basic issue of misclassification where a scatterplot of
true grades for each block are plotted against the corresponding predicted grades.

Figure 1: Misclassification in grade control. Source: (Rossi & Deutsch, 2014).

Common approaches include the classical nearest neighbour method where grades of the closest
blasthole sample are assigned directly to a block model, and the inverse distance weighting estimation
method which assign grades by calculating the weight of each sample that is inversely proportional to the
distance of the estimation location (Ortiz, 2020). Geostatistical methods such as kriging were introduced
in the early 1950s. Kriging minimizes the estimation variance under certain conditions (Rossi & Deutsch,
2014). Conditional simulation methods have been the most advocated technique in recent years to predict
grades and their uncertainty, thus addressing some of the shortcomings of kriging and the classical
methods. This method simulates grades at given locations and can be combined with different
optimization algorithms where the material types (ore or waste) are evaluated against all simulated
realizations and the optimum destination or material type at each location determined (Vasylchuk &
Deutsch, 2018).

In this paper (Vasylchuk & Deutsch, 2018), the authors review four grade interpolation methods:

1. Nearest neighbour polygonal method;

2. Inverse distance weighting method;
3. Ordinary kriging; and
4. Simulation-based methods.

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 60

The paper also discusses the performance of two novel machine learning (ML) algorithms that have been
used recently in grade control: Elliptical Radial Basis Function Network (ERBFN) and Support Vector
Regression (SVR). The results from these methods are analysed with the help of the case studies presented
to show which method is most efficient for grade control.

In the next sections, grade control methods are reviewed and critically assessed in a simple theoretical
framework. Results obtained in the literature from numerical experiments conducted to compare the
effectiveness of simulation versus different estimation methods, as well as results on a comparative case
study at the Carmen de Andacollo copper mine (Chile) are presented. Summary and conclusions follow.

2. Methods for grade prediction

Grade prediction is probably the most crucial aspect of grade control because it forms the basis for
selecting ore and waste zones. The main purpose of estimation is to predict the grade of the variable at
unsampled locations in the block model which is the premise for classification. Samples taken from blast
holes are analyzed from which a quantitative model of the ore body is constructed by interpolating and
extrapolating between these samples to account for the grade in areas that were not sampled. This is with
the assumption that all the sample locations and the unsampled location belong to the same domain.
There are various methods developed for performing grade prediction in grade control, however this
paper discusses the most common ones in the industry.

2.1. Nearest Neighbour (NN) estimation method

The nearest neighbor method is one of the simplest approaches to grade estimation. As a variant of the
polygonal method, the nearest neighbor assigns the grade of the closest blast hole sample to the entire
unsampled block (Vasylchuk, 2019). Since the weight of each sample is an important factor in estimating,
NN method determines the weight of the samples by assigning all the weight to the closest sample and
every other sample gets a weight of zero. In relation to the cut-off grade, the NN method regards the
estimation location as ore if the grade of the closest sample is larger than the cut-off grade. The value at
the unsampled location is then calculated as shown in the equation below:
∗ (𝑢 ) 𝑁𝑁 𝑁𝑁𝑛
𝑧𝑁𝑁 𝑜 = 𝜆𝑜 + ∑𝑖=1 𝜆𝑖 𝑧(𝑢𝑖 ) (1)

1 𝑖𝑓 𝑢𝑖 𝑖𝑠 𝑐𝑙𝑜𝑠𝑒𝑠𝑡 𝑡𝑜 𝑢𝑜
𝜆𝑁𝑁
𝑖 = { } 𝑖 = 1, … , 𝑛
0 𝑜𝑡ℎ𝑒𝑟𝑤𝑖𝑠𝑒
𝜆𝑁𝑁
0 =0
∗ (𝑈 ) 𝑁𝑁
where 𝑧𝑁𝑁 𝑜 is the nearest neighbor estimate at the unsampled location (𝑢𝑜 ), 𝜆 are the nearest
neighbor weights of the samples, and 𝑧(𝑢𝑖 ) is the known grade of the samples for which 𝑖 = 1, … , 𝑛 are
the total number of samples.

One main strength of the nearest neighbour method is that it does not smooth estimated values
(Kapageridis, 2014). However, study has shown that this method is not particularly accurate and cannot
be trusted because it neither takes into account the spatial continuity of the grade nor the redundancy in
the information (Ortiz, 2020). The discrepancies in this method are known to be larger that other
estimators, and for many deposits that have positively skewed distributions, significant errors in the
estimate occur in the individual blocks leading to proclivity to overestimate the average grade and

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 61

underestimate tonnage above cut-off (Rossi & Deutsch, 2014). Nonetheless the nearest neighbour
method can be used as a checking tool.

2.2. Inverse Distance Weighting (IDW) estimation method

The inverse distance weighting technique is an enhancement of the classical polygonal method, and is
most suitable for uniform orebodies (Abuntori et al., 2021). This method is used to estimate grade values
using several nearby blasthole sample grades to obtain a weighted average for each block as shown in
Figure 2. The calculation of the estimate is shown below:
𝑛
∗
𝑧𝐼𝐷𝑊 (𝑢𝑜 ) = 𝜆𝐼𝐷𝑊
0 + ∑𝑖=1 𝜆𝐼𝐷𝑊
𝑖 𝑧(𝑢𝑖 ) 𝑖 = 1, … , 𝑛 (2)
∗ (𝑢𝑜 ) is the inverse distance estimate at the unsampled location (𝑢𝑜 ), 𝜆𝐼𝐷𝑊 is the inverse
where 𝑧𝐼𝐷𝑊
distance weight assigned to each known sample and 𝑧(𝑢𝑖 ) is the grade of each known sample for which
𝑖 = 0, … , 𝑛 are the total number of samples. In this scenario, the weights assigned to each sample are
inversely proportional to the distance from estimation location and are calculated as shown in equation
3 below. Each sample is weighted based on its proximity to the location to be estimated.
1⁄ 𝑤)
(𝑐+𝑑𝑖𝑜
𝜆𝐼𝐷𝑊
𝑖 = 𝑛 𝑖 = 1, … , 𝑛 (3)
∑ 1⁄ 𝑤)
(𝑐+𝑑𝑖𝑜
𝑖=1

𝜆𝐼𝐷𝑊
0 =0

for which 𝜆𝐼𝐷𝑊

𝑖 is the inverse distance weight of sample 𝑖, 𝑑𝑖𝑜 is the distance between the estimation
location and sample 𝑖, 𝑤 is the inverse distance weighting power and 𝑐 is a small constant for numerical
stability or computational reasons. When the weighting power approaches zero, the weights become
similar and is calculated as the arithmetic average of the samples (Abzalov, 2016). On the other hand, a
larger weighting power assigns all the weight to the closest sample making the inverse distance weight
similar to the result of the polygonal nearest neighbour method (Ortiz, 2020). In practice, the most
frequently used weighting power is 2, however powers of 1 and 3 are also used for estimation. Even
though the inverse distance weighting method provides better estimates than the nearest neighbor
method, it does not account for the details of the data configuration or the varying anisotropy at different
scales (Vasylchuk, 2019).

𝑧(𝑢1 )
𝑑1 𝑧(𝑢2 )
𝑑2
𝑧∗
𝑑3 𝑧(𝑢3 )
𝑑4

𝑧(𝑢4 )

Figure 2: Inverse Distance Weighting method.

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 62

 2.3. Kriging-based estimation method
Geostatistics, developed in the early 1950s, forms the basis for kriging (Krige, 1951). The geostatistical
concept provides the platform for describing and modelling the spatial continuities of the regionalised
variables (in this case the grade values) and allows incorporation of the continuity factors into the
regression techniques used for the spatial predictions (Abzalov, 2016). Kriging is a collection of generalized
linear regression techniques based on calculating optimal weights that minimize the expected error
variance or the estimation variance (Ortiz, 2020). It produces an estimate that is a weighted linear
combination of the data, minimizes the estimation error, hence, it called the Best Linear Unbiased
Estimator (BLUE).

Kriging-based grade control came to light in open pit mines during the 1980s (Deutsch et al., 2000).
Different types of kriging algorithms have been used in grade control, but most commonly ordinary kriging
(OK) particularly in gold mines in Northern Nevada (Rossi & Deutsch, 2014). Other types of kriging have
also been applied such as the indicator kriging and not too popular fuzzy kriging (González, 2012).
Although not very common, simple kriging (SK) has also been used to estimate but more often used as a
checking tool.

Ordinary kriging
Ordinary kriging (OK) is a robust estimator which assumes that the local mean is unknown (unlike SK which
assumes a known mean), but constant within the estimation neighborhood based on the quasi second
order stationarity assumption (Ortiz, 2020). To guarantee global unbiasedness, OK constrains the sum of
the weights to be 1.0, and as a result the mean does not need to be known (Abuntori et al., 2021). The
weights used in kriging directly depend on the choice of a variogram model for the data set. The variogram
model, which is a prerequisite, enables the kriging algorithm to obtain insight on the anisotropy in the
grade distribution. The ordinary kriging estimate is summarized in the equation below.
∗ (𝑢 ) 𝑂𝐾 𝑛
𝑂𝑟𝑑𝑖𝑛𝑎𝑟𝑦 𝑘𝑟𝑖𝑔𝑖𝑛𝑔 𝑒𝑠𝑡𝑖𝑚𝑎𝑡𝑜𝑟, 𝑍𝑂𝐾 𝑜 = ∑𝑖=1 𝜆𝑖 𝑍(𝑢𝑖 ) (4)
∗ (𝑢 ) 𝑂𝐾
where 𝑍𝑂𝐾 𝑜 is the ordinary kriging value at the unsampled location (𝑢𝑜 ), 𝜆𝑖 the weight assigned to
each known sample (𝑢𝑖 ), and 𝑍(𝑢𝑖 ) is the grade of each known sample 𝑖 for which 𝑖 = 1, … , 𝑛 are the
total number of samples. The kriging variance which measures the quality of the estimation is given by
2 (𝑢 ) 2 𝑂𝐾 𝑛
𝑂𝑟𝑑𝑖𝑛𝑎𝑟𝑦 𝑘𝑟𝑖𝑔𝑖𝑛𝑔 𝑣𝑎𝑟𝑖𝑎𝑛𝑐𝑒, 𝜎𝑂𝐾 𝑜 = 𝜎0 − ∑𝑖=1 𝜆𝑖 𝐶𝑖𝑜 − 𝜇 (5)
2 (𝑢 ) 2
where 𝜎𝑂𝐾 𝑜 is the kriging variance at the estimation location (𝑢𝑜 ), 𝜎0 is the variance of the
distribution, 𝜆𝑂𝐾
𝑖 is the weight assigned to the known sample 𝑖, 𝐶𝑖𝑜 is the covariance between sample 𝑖
and the estimation location (𝑢𝑜 ), and 𝜇 is the Lagrange multiplier which is an additional parameter to
help in optimality.

There are other variants of the kriging method which have proven successful in many operations such as
the Breakeven Indicator Method (BEI) (Vasylchuk, 2016). The BEI grade control method is a blend of both
indicator and grade kriging. It uses an ore/waste indicator variable to predict the probability of ore
occurrence at a given location (Rossi & Deutsch, 2014). The indicator variable is then used to define ore
or waste probability of the estimated value based on the grade of the blast holes, and the expected
revenue is determined. This method was used in copper-molybdenum Ujina open-pit in Chile, together
with and the classical inverse distance weighting method and the results were compared to a reference

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 63

model. The BEI showed a relatively better performance than IDW and a summary of the results is shown
in subsequent sections.

In summary, kriging provides good frameworks for predicting grades that are locally accurate estimates,
however, the premise of estimating based on the minimization of estimation variance is not optimal for
grade control (Srivastava, 1987). (Rossi & Deutsch, 2014) recounts that kriging has been only slightly more
successful at grade control compared to the other classical methods because of the inherent smoothing
and the inability to quantify the spatial uncertainty as shown in Figure 3.

Figure 3: Smoothing effect of kriging.

2.4. Simulation-based method

The use of simulation as a predicting tool has been a cutting-edge method in grade control. The
commonest simulation method used to model the realistic variability of a deposit is the Sequential
Gaussian Simulation (SGS) (Vasylchuk & Deutsch, 2018). This method assigns grade to blocks and also
takes into account the uncertainty in the grade distribution that can be later used for assessing economic
consequences of grade control decisions, a feature lacked by the traditional estimation methods
(Vasylchuk, 2019). The grade assignment process involves a series of steps such as data declustering,
normal score transformation of declustered data, simulating a value from the conditional distribution and
back-transforming simulated values. The simulation process returns a range of probable values from a
conditional cumulative distribution function (cdf) as shown in the equation below:

𝐹(𝑢; 𝑧|(𝑛)) = 𝑃𝑟𝑜𝑏{𝑍(𝑢) ≤ 𝑧|(𝑛)} (6)

where 𝐹(𝑢; 𝑧|(𝑛)) is the cumulative frequency distribution curve, 𝑍(𝑢) accounts for the uncertainty in
the unknown true value, and (𝑛) represents local conditioning blast hole data within the specific
neighborhood of location (𝑢).

In order to evaluate which realization will produce the optimum classification of materials, taking into
consideration its economic impact on the operation, simulation-based methods incorporate certain
optimization algorithms such as minimum loss and maximum profit functions (Deutsch et al., 2000;
Dimitrakopoulos & Godoy, 2014; Vasylchuk, 2016). Moreover, the economic consequences of sending ore
to the waste dump is different from sending waste to the mill, hence an optimal selection criterion is
needed to account for these asymmetric economic impacts through the optimization algorithms that

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 64

simulation provides. The application of the economic classification functions for grade control does not
only account for the penalties of each decision, but also provide essential information for non-linear metal
recoveries or any other geo-metallurgical attribute of interest by adopting the economic functions
(Wambeke & Benndorf, 2017).

The minimum expected loss method consists of computing the expected loss associated with each
classification and selecting the classification for which the expected loss is minimal. Several mathematical
expressions have been provided by different authors for the ‘minimum loss’ function, but for the purposes
of subsequent comparison, a simplified version of the loss function by (Vasylchuk, 2016) is presented, as
shown in equation 7.
0, 𝑓𝑜𝑟 𝑎 𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑑𝑒𝑐𝑖𝑠𝑖𝑜𝑛
𝑇ℎ𝑒 𝑙𝑜𝑠𝑠 𝑓𝑢𝑛𝑐𝑡𝑖𝑜𝑛, 𝑔(𝑢; 𝑧, 𝑧𝑐 ) = {(𝑧(𝑢) − 𝑧𝑐 ) × 𝑏1 , 𝑓𝑜𝑟 𝑎𝑛 𝑖𝑛𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑤𝑎𝑠𝑡𝑒 𝑑𝑒𝑐𝑖𝑠𝑖𝑜𝑛
(𝑧𝑐 − 𝑧(𝑢)) × 𝑏2 , 𝑓𝑜𝑟 𝑎𝑛 𝑖𝑛𝑐𝑜𝑟𝑟𝑒𝑐𝑡 𝑜𝑟𝑒 𝑑𝑒𝑐𝑖𝑠𝑖𝑜𝑛

Hence, 𝑡ℎ𝑒 𝐸𝑥𝑝𝑒𝑐𝑡𝑒𝑑 𝑙𝑜𝑠𝑠 𝑑𝑒𝑐𝑖𝑠𝑖𝑜𝑛 = 𝐸[𝑔(𝑢; 𝑍, 𝑧𝑐 )] (7)

where 𝑔(𝑢; 𝑍, 𝑧𝑐 ) represents the loss function, 𝐸[𝑔(𝑢; 𝑍, 𝑧𝑐 )] is the expected loss over multiple
realizations at location (𝑢), 𝑧𝑐 is the cut-off grade, 𝑧(𝑢) is the simulated grade value at location (𝑢), 𝑏1
and 𝑏2 are the penalty coefficients for underestimating and overestimating respectively. For instance, a
block will be selected as ore if the expected loss for mining the block as ore is less than the expected loss
for mining it as waste.

Similarly, for maximum profit function, a block will be selected as ore if the expected profit for mining the
block as ore exceeds the expected profit for mining it as waste, and vice versa. More information on
concept of minimum expected loss or maximum profit as a basis for classification decisions can be found
in (Rossi & Deutsch, 2014; Vasylchuk, 2016; Verly, 2005; Wambeke & Benndorf, 2017).

2.5. Machine Learning (ML) method

Technological advancements in recent years have enabled computers to process large amounts of
information within the shortest possible time. Machine learning algorithms (MLA) are a collection of
advanced statistical methods empowered by high-level computers to provide flexibility and simplicity
when integrating and recognizing complicated patterns in data, which is a difficult task with linear
geostatistical workflows. Machine learning is already incorporated to solve Earth Sciences problems
(Deutsch et al., 2016) but is not fully integrated into the geometallurgical workflows that model and
optimize the processes leading to the extraction and recovery of minerals and metals (Ortiz, 2019). As far
as this review is concerned, the application of machine learning methods for grade assignment and
classification in grade control is a novel enterprise with not too many applications. ML algorithms such as
Artificial Neural Networks have been used for mineral resource estimation (Abuntori et al., 2021) but have
not been explored in grade control. The application of ML is to enhance the accuracy of predicted grade
values, and to make better decisions concerning the classification of mined material.

A recent work by (Da Silva et al., 2020) demonstrated the application of ensemble ML methods for grade
control in the Carmen de Andacollo copper mine in Chile. Two algorithms were used in their work:
Elliptical Radial Basis Function Network (ERBFN) and Support Vector Regression (SVR). The two ML
algorithms are trained which was preceded by the tuning of their respective hyperparameters. For
instance, the number of nodes that constitute a hidden layer in the network was defined for ERBFN. Once

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 65

the nodes are defined, ERBFN algorithm trains the data and assigns a new networking system to each
node to make grade predictions. The different predictions obtained are averaged to form a trend model
which is further used as a secondary variable to assess the consequences of each grade control decision
on an intrinsic collocated cokriging framework. Just as kriging, variogram models were generated and
search parameters defined.

In the case of SVR, the data is divided into a training, validation, and test dataset. Different reference
models are defined from which predictions are generated for the validation set. The predicted values and
the blasthole values are then fed into a trained meta model from these pairs of values, and then final
grade predictions are made. Just like ERBFN, the SVR algorithm optimizes the best scenario for ore and
waste by using the final predicted model in a collocated cokriging framework. Finally, based on a break-
even cut-off grade defined by the mine operations, the destination of each material is determined. Details
of this work can be found in (Da Silva et al., 2020).

3. Results and Discussions

The results presented here are case studies reviewed from (Da Silva et al., 2020; Rossi & Deutsch, 2014;
Vasylchuk, 2016) which demonstrate the performance of grade prediction methods. The first scenario is
a study in a copper-molybdenum Ujina open pit mine in Northern Chile where the outcomes of the Break-
even Indicator (BEI) method, and the inverse distance weighting (IDW) method are compared to a
reference model in terms of their ability to classify and provide destination for materials. The results show
that BEI is superior to IDW in that it produced results closer to the reference model. The study further
stated that the simulation based approached used here produced results similar yet slightly better than
the BEI. Meanwhile, Ordinary kriging (OK) produced a marginally inferior result. Only results for BEI and
IDW are presented here with respect to the reference model for tonnages and total Cu grade for different
destinations. Proximity of a value to 1.0 indicates better performance of the method. A factor greater
than 1 implies overestimation with respect to the reference model. Details of this work can be found in
(Rossi & Deutsch, 2014).
Table 1: Performance of IDW and BEI models with respect to an SGS reference model. Source: (Rossi & Deutsch, 2014).

Destination code Tonnage w.r.t reference TCu Grade w.r.t reference

IDW BEI IDW BEI
SAL 1.10 1.10 0.91 0.92
SME 1.16 1.09 1.06 1.00
SBA 0.18 0.45 1.15 1.01
SMR 0.50 0.43 1.36 1.01
SAS 0.55 0.87 1.02 0.95
OXA 1.29 1.13 0.85 0.93
OXB 1.16 1.98 1.08 0.98
OXL 0.44 1.49 1.54 1.41
MIX 0.52 0.71 0.90 0.78
TOTAL 1.16 1.11 0.84 0.89

In the second scenario, a numerical experiment was conducted to evaluate the effectiveness of simulation
versus different estimation methods based on the losses incurred as derived from the expected loss

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 66

function. The penalty coefficients for underestimating and overestimating, 𝑏1 and 𝑏2 respectively are
given as a ratio. A ratio of 1:1 means the consequences of under or overestimating are equal. If the ratio
is asymmetric (example 1:2), it means that the penalty for overestimating is higher. Only a few ratios are
shown here for illustrative purposes. The results show that the simulation-based method incurred the
least losses than the other methods both on average and on individual penalty considerations.
Table 2: Incurred losses of grade control methods. Source: (Vasylchuk & Deutsch, 2018).

GC 𝒃𝟏 : 𝒃𝟐
Methods 2.65:1 2:1 1.3:1 1:1 1:1.3 1:2 1:2.65 Average
NN 1208.4 433.3 257.0 206.8 246.6 387.3 1006.0 454.0
ID 900.6 333.3 204.2 166.4 203.9 331.8 894.2 369.3
OK 877.7 323.7 197.6 160.6 196.4 318.5 855.1 359.8
SK 861.9 316.5 192.4 156.0 190.3 307.2 820.9 349.8
Simulation 338.7 215.6 186.7 155.8 186.3 249.7 355.0 224.7

The last scenario involves the integration of machine learning techniques into grade control. The two ML
algorithms used (ERBFN and SVR) were compared to other commonly used methods in the industry with
respect to how their grade assignment led to the classification of the materials. The methods were Inverse
distance (ID), ordinary kriging (OK), and an intelligent grade control (IGC) based on multivariate simulation.
From the study, it was observed that ERBFN and SVR, in conjunction with the collocated cokriging
framework outperformed traditional estimation methods and the simulation-based method. The efficient
grade prediction process led to better classification and better prediction of material destinations. In
Figure 4 below, a five-fold mean square error (MSER) validation is used to measure the performance of
each model over ten blast holes considered. The red points represent the overall MSER for each method.
From the study, it was observed that the inverse distance estimation (ID) obtained the highest MSER of
0.00888, while that of ERBFN and SVR obtained an MSER of 0.007 and 0.0075 respectively.

Finally, the extent of misclassification by the methods were evaluated and the study revealed that ERBFN
with collocated cokriging (CCok) obtained the best performance in reducing misclassification. It obtained
a reduction in misclassified material of 12% when compared to ID; 5.4% less than OK, 5.7% relative to IGC
and 4.3% relative to CCok with SVR. Even though ERBFN-CCok obtained better results that SVR-CCok, the
latter also obtained a reduction of 8% in misclassified material relative to ID and 1.12% and 1.16% to OK
and IGC respectively as shown in Figure 5.

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 67

Figure 4: Mean squared error obtained from a 5-fold cross validation for each method applied (source: (Da Silva et
al., 2020).

Figure 5: Total number of misclassified blocks recorded by the grade control methods (source: (Da Silva et al., 2020).

4. Conclusions
Grade control methods for classifying materials have evolved. The methods reviewed in this paper have
shown that classical estimation methods can no longer be relied on in making grade control decisions. The
smoothing effect and other limitations of kriging methods are a huge concern for grade control and hence
do not make it an optimal method. Simulation based methods have provided good results which still make
them a cutting-edge tool in grade control and is still used by most mining companies today. However,
machine learning methods may become the future of grade control due to its high performance in grade
prediction and significantly reducing misclassification, which is the goal of grade control.

5. Acknowledgments
We acknowledge the support of the Natural Sciences and Engineering Research Council of Canada
(NSERC), funding reference number RGPIN-2017-04200 and RGPAS-2017-507956.

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 68

6. References
Abuntori, C. A., Al-Hassan, S., & Mireku-Gyimah, D. (2021). Assessment of Ore Grade Estimation Methods
for Structurally Controlled Vein Deposits—A Review. Ghana Mining Journal, 21(1), 31–44.
https://doi.org/10.4314/gm.v21i1.4

Abzalov, M. (2016). Applied Mining Geology (Vol. 12). Springer International Publishing.
https://doi.org/10.1007/978-3-319-39264-6

Da Silva, C. Z., Nisenson, J., & Boisvert, J. (2020). Grade control with Ensembled ML: A comparative case
study at the Carmen de Andacollo copper mine [Annual report].

Deutsch, C. V., Magri, E., & Norrena, K. (2000). Optimal Grade Control Using Geostatistics and Economics.
Society for Mining, Metallurgy, and Exploraction Inc., 308.

Deutsch, J. L., Palmer, K., Deutsch, C. V., Szymanski, J., & Etsell, T. H. (2016). Spatial Modeling of
Geometallurgical Properties: Techniques and a Case Study. Natural Resources Research, 25(2), 161–
181. https://doi.org/10.1007/s11053-015-9276-x

Dimitrakopoulos, R., & Godoy, M. (2014). Grade control based on economic ore/waste classification
functions and stochastic simulations: Examples, comparisons and applications. Mining Technology,
123(2), 90–106. https://doi.org/10.1179/1743286314Y.0000000062

Giraldo, R., Herrera, L., & Leiva, V. (2020). Cokriging Prediction Using as Secondary Variable a Functional
Random Field with Application in Environmental Pollution. Mathematics, 8(8), 1305.
https://doi.org/10.3390/math8081305

Giraldo, R., Mateu, J., & Delicado, P. (2017). Cokriging and multivariate kriging methods based on data of
a functional random field. 10(2), 30.

Gräler, B. (n.d.). Cokriging and indicator kriging. 21.

Kapageridis, I. K. (2014). APPLICATION OF ARTIFICIAL NEURAL NETWORK SYSTEMS TO GRADE ESTIMATION

FROM EXPLORATION DATA. 268.

Krige, D. G. (1951). A Statistical Approaches to Some Basic Mine Valuation Problems on the
Witwatersrand. Journal of the Chemical, Metallurgical and Mining Society of South Africa, 52, 119–
139.

Moreira, G. de C., Modena, R. C. C., Costa, J. F. C. L., & Marques, D. M. (2021). A workflow for defining
geological domains using machine learning and geostatistics. Tecnologia Em Metalurgia, Materiais e
Mineração, 18, e2472. https://doi.org/10.4322/2176-1523.20212472

Ortiz, J. M. (2019). Geometallurgical modeling framework, Predictive Geometallurgy and Geostatistics

Lab, Queen’s University, Annual Report 2019, paper 2019-01, 6-16.

Ortiz, J. M. (2020). Advanced Geostatistics—Estimation. Queen’s University, 29.

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 69

Rossi, M. E., & Deutsch, C. V. (2014). Mineral Resource Estimation. Springer Netherlands.
https://doi.org/10.1007/978-1-4020-5717-5

Srivastava, R. M. (1987). Minimum variance or maximum profitability. CIMM, 80(901), 63–68.

Vasylchuk, Y. V. (2016). Integrated System for Improved Grade Control in Open Pit Mines. Master of
Science in Mining Engineering thesis, University of Alberta, 150.

Vasylchuk, Y. V. (2019). Advanced Grade Control with Multivariate Geostatistics, Blast Movement
Modeling, and Optimized Dig Limits. Doctor of Philosophy thesis, University of Alberta, 236.

Vasylchuk, Y. V., & Deutsch, C. V. (2018). Improved grade control in open pit mines. Mining Technology,
127(2), 84–91. https://doi.org/10.1080/14749009.2017.1363991

Verly, G. (2005). Grade Control Classification of Ore and Waste: A Critical Review of Estimation and
Simulation Based Procedures. Mathematical Geology, 37(5), 451–475.
https://doi.org/10.1007/s11004-005-6660-9

Wambeke, T., & Benndorf, J. (2017). A Simulation-Based Geostatistical Approach to Real-Time

Reconciliation of the Grade Control Model. Mathematical Geosciences, 49(1), 1–37.
https://doi.org/10.1007/s11004-016-9658-6

© Predictive Geometallurgy and Geostatistics Lab, Queen’s University 70