Natural Resources Research, Vol. 33, No.

1, February 2024 (Ó 2023)

https://doi.org/10.1007/s11053-023-10280-7

Original Paper

Spatial Interpolation Using Machine Learning: From Pat-

terns and Regularities to Block Models

Glen T. Nwaila ,1,6 Steven E. Zhang,2 Julie E. Bourdeau,2 Hartwig E. Frimmel,3,4 and
Yousef Ghorbani5

Received 14 May 2023; accepted 26 October 2023

Published online: 25 November 2023

In geospatial data interpolation, as in mapping, mineral resource estimation, modeling and

numerical modeling in geosciences, kriging has been a central technique since the advent of
geostatistics. Here, we introduce a new method for spatial interpolation in 2D and 3D using
a block discretization technique (i.e., microblocking) using purely machine-learning algo-
rithms and workflow design. This paper addresses the challenges of modeling spatial patterns
and regularities in nature, and how different approaches have been used to cope with these
challenges. We specifically explore the advantages and drawbacks of kriging while high-
lighting the long and complex sequence of procedures associated with block kriging. We
argue that machine-learning techniques offer opportunities to simplify and streamline the
process of mapping and mineral resource estimation, especially in cases of strong spatial
relationships between sample location and resource concentration. To test the new method,
synthetic 2D and 3D data were used for both 2D block modeling and geometallurgical
modeling of a synthetic porphyry Cu deposit. The synthetic porphyry Cu data were very
useful in validating the performance of the proposed microblocking technique as we were
able to reproduce known values at unsampled locations. Our proposed method delivers the
benefits of a machine learning-based block modeling approach, which includes its simplicity
(a minimum of 2 hyperparameters), speed and familiarity to data scientists. This enables
data scientists working on spatial data to employ workflows familiar to their training, to
tackle problems that were previously solely in the domain of geoscience. In exchange, we
expect that our method will be a gateway to attract more data scientist to become geodata
scientists, benefitting the modern data-driven mineral value chain.
KEY WORDS: Geostatistics, Block model, Kriging, Mineral resources estimation, Machine learning,
Patterns.

1
Wits Mining Institute, University of the Witwatersrand, 1 Jan
Smuts Ave., Johannesburg 2000, South Africa. INTRODUCTION
2
Geological Survey of Canada, 601 Booth Street, Ottawa, ON
K1A 0E8, Canada. Patterns and regularities are common phe-
3
Institute of Geography and Geology, Department of Geody- nomena in nature and represent a state of order and
namics and Geomaterials Research, University of Würzburg, Am
structure, and rational principles that govern nature
Hubland, 97074 Würzburg, Germany.
4
Department of Geological Sciences, University of Cape Town, (Good, 1983; Washburn et al., 1988; Dennett, 1991;
Rondebosch 7701, South Africa. Goertzel, 2006; Uddin & Hamiduzzaman, 2009;
5
School of Chemistry, University of Lincoln, Joseph Banks Lab- Kuipers, 2001; Steiner, 2009). They have been
oratories, Green Lane, Lincoln, Lincolnshire LN6 7DL, UK. studied in a variety of fields, such as physics, math-
6
To whom correspondence should be addressed; e-mail: glen.n-
ematics, statistics, chemistry, biology, philosophy
[email protected]

129
1520-7439/24/0200-0129/0 Ó 2023 The Author(s)
130 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

and geosciences. In geosciences, spatial patterns are drawback of ML, we demonstrate that this property
evident in the repetitive and consistent characteris- could be exploited, in combination of a re-exami-
tics of rock types, mineral deposits and their for- nation of the assumption of support punctuality to
mation processes (Groves et al., 2005). Spatial perform a rigorous change-of-support. We compare
characteristics are of paramount importance in and contrast the results using ML and typical geo-
exploring for, and comprehending, the genesis of statistical modeling using a large ensemble of syn-
mineral deposits and can be leveraged to improve thetically generated datasets. We explore the
spatial data modeling from mapping to mineral re- differences between the traditional geostatistical
source estimation. The study of spatial patterns has approach and our proposed method in a variety of
been a focus of research in geosciences for hundreds scenarios, using a combination of quantitative and
of years and has led to the development of theories qualitative comparisons through modulating a few
that propose that the repetitive and consistent key parameters: (1) sampling type – regular and
characteristics of mineral deposits can be used to biased random sampling; (2) sampling rate (both
locate, decipher ore-forming processes and facilitate regular and biased random sampling); (3) strength of
mineral resource estimations (Carranza, 2009). spatial correlation (through controlling the nugget
In geosciences, spatial patterns are important to effect); and (4) anisotropy. The results indicate that
both knowledge and data-driven inquiries. Methods qualitatively and quantitatively, our method pro-
to model and study spatial patterns have resulted in duces results that are competitive with block kriging
the birth and development of many disciplines, such across a wide range of conditions. A key benefit of
as geostatistics and remote sensing. Compared to our approach is that it requires a substantially re-
traditional geostatistics, machine learning (ML), as a duced set of parameters to tune relative to block
field, is more recent and presents pathways toward kriging to produce similar results, and their tuning
leveraging data that describe mineral deposits and process is fully metric-based and automated, capi-
their signatures (Samson, 2020; Dumakor-Dupey & talizing on the ML framework, which offers a
Arya, 2021). ML algorithms have been employed in workflow of inter-compatible and automat-
the study and exploration of mineral deposits to able methods (e.g., cross-validation and model
identify and model patterns and regularities that are selection). Other benefits include: (1) freely avail-
unobvious (Srinivasan & Fisher, 1995; Galetakis able and mostly open-source libraries that are tuned
et al., 2022; Mery & Marcotte, 2022). This has been for high performance computation across a range of
demonstrated in studies such as applying ML algo- platforms; and (2) lower barrier of entry into spatial
rithms to satellite imagery to locate and study min- modeling for practitioners of artificial intelligence,
eral deposits, and to improve mineral exploration ML and data science by reformulating block mod-
(Maxwell et al., 2018; Cevik et al., 2021; Diaz-Gon- eling into ML workflows. Hence, in addition to sat-
zalez et al., 2022; Liu et al., 2022; Nwaila et al., isfactory performance characteristics, our method is
2022). Particularly related to this study is the use of more reproducible, less subjective and more acces-
ML to perform geodomain boundary delineation sible. Thereafter, we deploy our method to a simu-
(Zhang et al., 2023), which is a task that is required lated porphyry Cu deposit (Garrido et al., 2018,
in the geostatistical treatment of mapping to re- 2020) to create three dimensional (3D) geometal-
source estimation. Similarly, ML algorithms that lurgical block models.
could model spatial patterns are a powerful tool to
improve or provide alternatives to geostatistical
spatial modeling. The rigor of resource estimation SYNOPTIC REVIEW OF GEOSTATISTICS
(e.g., a need for reconciliation) implies that differ- AND COMMON CHALLENGES
ences in spatial modeling are the most impactful and
appreciable at this scale, although any benefits of a Geostatistics is a well-developed domain of
new approach would apply across all spatial scales. geosciences with an extensive and impressive his-
In this paper, we propose a new method to tory, as well as a large variety of known applications
perform spatial interpolation using ML algorithms in (primarily in solid-earth science and related fields)
a manner similar to that of geostatistical block that include mapping, resource modeling, spatial
modeling. A key recognition in our method is that prediction, and interpolation and change of resolu-
ML algorithms are generally unable to perform a tion (Krige, 1997; Ortiz & Emery, 2006; Talebi et al.,
change-of-support. Although this seems like a 2019). Its fundamental premise is that many natu-
Spatial Interpolation Using Machine Learning: from Patterns 131

rally occurring spatial phenomena that are not sub- A particular solution to the change-of-support
jected to a high driving rate of mixing and transport problem for the purpose of block modeling is block
processes exhibit a spatial variability (essentially kriging, although this concept is also relied upon
pseudo-equilibrium spatial distributions; Matheron, across the broader geostatistics. Within the domain
1967). These spatial patterns can be quantified to of geostatistics and for the purpose of resource
produce models of the variability structure, which, in estimation, this is the most commonly used method.
turn, becomes useful for estimation, mapping, gen- Although block models are often visualized as
eral interpolation and other purposes (Isaaks & quantized (discretized) representations of a physical
Srivastava, 1989; Isaaks, 2005). In other words, the phenomenon (e.g., a portion of an orebody), a range
statistics of sampled data of many solid-earth pro- of discretization exists such that the perception of
cesses are expected to exhibit a general spatial discretization is not always apparent (Vann et al.,
variability, and therefore, correlatability. A partic- 2003). For example, in the use of geostatistical
ular task that is commonly performed using geo- interpolation methods to create regional maps from
statistics is block modeling. Although block geochemical concentrations, the effect of quantiza-
modeling is often used for resource estimation tion can be either masked by large pixel sizes or
within the context of the mineral value chain, it is post-hoc smoothing, and hence, block-level quanti-
implicitly used for other purposes as well. This is zation is not always obvious (Abzalov & Hum-
because block modeling is a solution to the Ôchange- phreys, 2002). For resource estimation, quantization
of-support problemÕ (Gelfand et al., 2001; Gotway & is usually visually obvious because absolute dis-
Young, 2002). The idea of a geostatistical support tances involved are typically much smaller than re-
refers to the ideal dimensionality of a geospatial gional maps and discretized blocks favor extraction
measurement, which could be 1D (a point sample), sequencing and reconciliation, for which smoothing
2D (a surface response) and 3D (a volumetric re- is also undesirable (Sarma, 2009). Block kriging is
sponse). Changing support refers to changing the therefore capable of creating 2D or 3D models of
representation of the geospatial phenomena from inferred reality at a range of quantization using
one type of support to another. Creating maps, for typically 1D data. This process is followed under
example, requires that a 1D support be changed into implicit or explicit assumptions: (1) that the 1D
a 2D or even 3D support to represent an areal or samples are infinitesimally small, such that their
volumetric distribution of some quantity. Change of internal structure is omitted at the scale of obser-
support is an unsolved problem in the general sense, vation (or effectively a 0-dimensional, punctual
because there is no unique and universally applica- support); and (2) that samples are representative of
ble solution, and various disciplines such as geo- spatial variability at the scale of observation. How-
statistics, produce a type of solution based on its ever, in practice, the correctness of the approxima-
capabilities and prior assumptions (Gelfand et al., tion of punctual support is a continuum and depends
2001; Gotway & Young, 2002). The change-of-sup- on the volume of the block versus the volume of the
port problem is commonly encountered in geo- sample, and in the extreme case that the two quan-
sciences due to two key characteristics: (1) native- tities are comparable, punctual support assumptions
support sampling, in which case data are collected at are clearly violated (Matheron, 1967; David, 1976).
the exact spatial resolution at which the natural While sample-points-to-block or block-to-block
phenomenon occurs, is often hindered by sampling kriging is the industry-preferred method of mineral
opportunity constraints and resource limitations; resource estimation, it is not without drawbacks. In
and (2) the necessity for scientific reduction and many cases, poor implementation of kriging or
associated analysis techniques has been a significant subjective configuration of kriging parameters have
driving factor in shaping data-engineering methods led to the under- or over-valuation of regionalized
(Cressie, 1990; Carvalho and Deutsch, 2017). Al- variables (e.g., resources, Krige, 1997; Isaaks, 2005).
though data-generation methods are currently One of the primary issues in point and block kriging
evolving, primarily driven by an evolution in guiding is the number of (usually manual) steps involved,
scientific philosophy (from reductionism to system which can result in a time-consuming and challeng-
considerations) brought forth by the change in the ing implementation process, even with some of the
purpose of geoscientific data (toward increasing use sub-tasks automated. Furthermore, point or block
of transdisciplinary techniques, such as artificial kriging often requires an expert-based tuning of
intelligence, ML and data science). various parameters, which is manual and can involve
132 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

subjectivity, in turn decreasing reproducibility. learning relationships at a different resolution).

However, geostatistical modeling is not irrepro- However, this view is over-simplified and ignores that
ducible. In practice, reproducibility is usually en- the punctuality of support is an approximation nec-
sured via thorough rigorous training, documentation essary for geostatistical methods and that no physical
of the parameters and methodology. Moreover, sample can, in the strictest sense, become punctual. In
advancements in geostatistical software and the effect, all physical samples are finite in volume and
establishment of guidelines for parameter calibra- this fact can be exploited to assist ML algorithms to
tion have contributed to standardizing and facilitat- perform a rigorous change-of-support. The manner in
ing the parameter tuning process. which this could be performed is not unique, but a
Despite substantial progress toward standardiz- straightforward method is to also exploit the lack of
ing resource estimation workflows, there is still no support awareness in ML algorithms. The process
universal protocol, and resource assessment remains a consists of: (1) determination of an approximate
discipline in which only a few practitioners are con- support size; (2) making an at-scale prediction of a
sidered ‘‘Competent and Qualified Persons1’’ due to grid, which is sized such that each grid spacing is
the technical and mathematical difficulties involved. roughly the volume of the support (microblocks); and
This is less of an issue for typical mapping uses of (3) down-sampling of microblocks to macroblocks to
geostatistics because the knowledge required to esti- achieve the desired block size (integer down-sampling
mate resources exceed merely spatial interpolation, incurs zero re-sampling error). Our proposed method
although rigorous geostatistics is an entire discipline recognizes properties of geospatial data beyond the
and is not typically within the training of modern data data-driven realm, because the representativity of the
talents (e.g., data scientists). This is a solvable prob- data in terms of sample volume and geostatistical
lem toward attracting data talents into geosciences, as considerations are clearly required and neither are
we seek to demonstrate with our proposed method. strictly transdisciplinary knowledge by origin. This is
The solution requires a geostatistical examination of also a general method, because any ML algorithm and
ML algorithms. ML algorithms are used by data sci- geospatial data could be leveraged in this manner to
entists but are incapable, in general, of a rigorous directly produce block models at any desirable reso-
change-of-support in the manner of geostatistics lution. The punctuality assumption of geostatistics is
(Veronesi & Schillaci, 2019). This is because ML transformed into a microblock assumption in our
algorithms (including many deep learning architec- approach – that the size of microblocks must be suf-
tures) are unable to constrain local spatial structures. ficiently small compared to the size of macroblocks. In
Hence, relationships between features and data labels the case where the macroblock size is much larger
generally do not capture changes in support, even if than the support volume, it is not necessary that the
spatial coordinates were used as features. Indeed, for microblocks are sized to roughly the support volume,
the activity of mineral prospectivity mapping, al- because averages over an area of a large ensemble of
though many ML and deep learning algorithms have predicted microblocks would still converge to the
been used, all evidence layers (e.g., geochemical and macroblock mean. In this manner, the approximation
geophysical data) are interpolated before modeling of punctuality carries over into the ML-based meth-
using non-ML approaches (e.g., Lawley et al., 2021, od, but the implications are different to those of
2022; Parsa et al., 2023). Hence, predictions made geostatistics. Our approach is independently con-
using ML algorithms occur almost exclusively in the ceived but is mechanistically similar to some imple-
same support as features in data. This unfortunately mentations of block kriging, which eschews the ML
means that in cases where input data points are con- framework for geostatistical algorithms and workflow
sidered punctual, but the output is intended for vol- designs.
umetric representation, ML is not expected to yield
rigorous results and would generally not be able to
change output resolution without retraining (e.g., re- DATA AND METHODS
1
A Competent Person is a minerals industry professional (i.e., Synthetic Data Generation
registered as a professional of appropriate membership class and
organisation including recognised professional organisations) with
enforceable disciplinary processes including the powers to sus- The synthetic data were generated using the
pend or expel a member. Rules and regulations vary across sequential Gaussian simulation function in GSLIB
countries. (Deutsch & Journel, 1992), via the GeoStatsPy
Spatial Interpolation Using Machine Learning: from Patterns 133

interface in Python (Pyrcz et al., 2021). The gener- Machine Learning-Based Block Modeling
ation process consisted of: (1) creating a parameter
grid that included the sample rate, nugget effect and Formulating the block modeling problem into a
anisotropy; (2) holding all parameters but one as ML task requires two key considerations. First,
experimental control and varying the single param- predictions must be made at data label-volumes that
eter as the experimental variable to produce a var- are substantially smaller than the desired block size,
iogram model; (3) generating a realization using such that they approximate the condition of a static
sequential Gaussian simulation; and (4) sampling support. This satisfies the constraint that almost all
the realization at a specific grid-sampling interval. ML algorithms are unable to perform dynamic
This procedure allowed us to perform a controlled changes-of-support. Second, spatial coordinates are
exploration of the parameter space by isolating the to be used as the sole features for spatial estimation
effects of each parameter on our proposed approach tasks. No model deployment would occur beyond
versus that of block kriging within the parameter sampled areas (hence no spatially transferred
space. The anisotropy parameter was defined as the learning) because the models would be used solely
ratio of the minor and major ranges, such that if the for interpolation. To satisfy the first condition, the
anisotropy parameter were set to 1, the major and desired block size must be re-discretized into smaller
minor ranges would be equal. In this context, Ôsam- blocks, which we call ÔmicroblocksÕ. After predicting
ple rateÕ refers to the density or frequency of data all microblocks, they are averaged (via spatial down-
points or samples collected in the spatial domain. It sampling) to produce a macroblock model, which is
represents how closely spaced the data points are specified at the desired block size to perform a
within the area being sampled. A higher sample rate change-of-support. This type of down-sampling is
corresponds to more closely spaced data points, known as a type of spatial signal compression
whereas a lower sample rate indicates more widely (Crochiere and Rabiner, 1983). Additional knowl-
spaced data points. Two types of sampling methods edge of signal processing is useful in our approach
were explored – regular sampling, which used a fixed but is not a critical component. The key signal pro-
spacing to designate sample sites, and biased ran- cessing consideration is that the down-sampling
dom sampling, which used a smaller combination of should ideally be of an integer factor to incur no re-
a fixed spacing and an additional number of sample gridding error.
sites that are randomly scattered. The nugget effect For microblock grid spacing, we adopted a lin-
was explored by varying the nugget parameter. ear spacing of 1/10 of that of the block model pro-
Modulating the nugget effect of the synthetically duced by the synthetic data generation process (e.g.,
generated data allowed us to capture physically each macroblock comprises 100 microblocks). This
realistic microscale variability and/or measurement is a heuristic setting in our case to satisfy the
error, which permitted us to understand the capa- approximation that the size of microblocks is much
bility of our proposed method with increasing nug- smaller than that of macroblocks, because there is
get effect. The entire parameter grid is given in no effective volume associated with synthetic data.
Table 1. The other key fixed parameters included Based on our synthetic data, there was no evidence
minimum and maximum grid values ( 3.0, 3.0); of increasing accuracy of macroblocks with finer
grid size (100 cells by 100 cells); grid spacing (10 microblocks. In general, documentation of actual
cells); sill (1.0); azimuth (135°); major range (800 data generation processes would indicate the sample
cells); and variogram model (spherical). As the grid volume (e.g., Armstrong & Champigny, 1989;
exploration was performed via single parameter Cressie, 1990; Annels, 1991). The microblocks were
sweeps, the default (fixed) nugget effect was 0.1 and then predicted using a ML algorithm and subse-
the anisotropy was 0.625 during the sample rate quently, macroblocks were produced by down-sam-
sweeps. During the nugget effect sweep, the fixed pling the microblocks by an integer factor of 10,
sample rate was a spacing of 30 cells and the ani- retrieving the original block configuration of the
sotropy was 0.625. Lastly, during the anisotropy synthetic data (100 cells by 100 cells grid size with 10
sweep, the fixed sample rate was a spacing of 30 cells cell blocks).
and the nugget effect was 0.1. This was repeated for In a strict technical sense, ML algorithms, as a
biased random sampling method with the sample whole, are not generally intended for spatial mod-
spacing replaced by its corresponding value (see eling, because they were originally designed for
Table 1) for the number of random samples. feature space characteristics that encompass essen-
134 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Table 1. Parameter ranges for the explored parameters and their interval construction methods

Parameter Parameter range Interval construction

Sample rate – regular sampling 1 per 23 cells to 40 cells 1 cell intervals

Sample rate – biased random sampling 625 to 1736 cells total Area of grid divided by the square of the regular sampling rate
Nugget 0 to 0.70 Intervals of 0.05 (out of a total sill of 1.00)
Anisotropy 0.1 to 1.0 20 evenly divided intervals

tially all types of data. Of the entire superset of data 2005; Kotsiantis et al., 2007) and Gaussian processes
characteristics, very few are typical of spatial data, (Rasmussen & Williams, 2006; Kotsiantis et al.,
which makes ML, in general, unspecific to spatial 2007). This is because neighbor-based methods use
learning tasks. For spatial data, some common averages of neighbors, which is similar to inverse
characteristics include: (1) generally low feature distance modeling, and the Gaussian process is a ML
space dimensionality (usually 2 to 3 dimensions in generalization of kriging, although without as many
non-temporal data); (2) explicit connotations of hyperparameters, and hence, somewhat less flexible.
spatial feature interactions, e.g., in the form of spa- Other methods include a variety of tree-based
tial (linear) correlation; (3) primarily linearly inde- methods, such as random forest and boosted meth-
pendent feature coordinates (e.g., the Cartesian ods (e.g., Ho, 1995; Breiman, 1996a, b; Freund &
coordinate system and the use of the Euclidean Schapire, 1997; Kotsiantis, 2014; Sagi & Rokach,
metric to measure distance); (4) low sample rate 2018). These methods are not suitable for block
(resulting in sparse samples) in the case of manual modeling in the traditional sense, because they do
data generation; and (5) potentially noisy data with a not include geometric characteristics in the feature
spatial noise distribution (e.g., high nugget effect). In space, and in the presence of sparse samples (which
particular, although characteristics (1), (2), (3) and is the case of most geoscientific sampling), they
(4) do not preclude a variety of ML algorithms, they create orthogonal decision boundaries in lower
ameliorate the advantage of many algorithms by number of dimensions and, therefore, result in cross-
design. For example, by design and usage, tree- hatched patterns in block models that are unrealistic
based methods and artificial neural networks are for our purpose, at least without post-processing
intended to better model data with high-dimen- (e.g., Zhang et al., 2023). Additionally, tree-based
sional, complex, and nonlinear feature interactions. methods excel at leveraging nonlinear feature
If they are used with trivial data, such as those that interactions in high-dimensional feature space,
would warrant a linear regression, then their algo- which is a benefit that would not be easily realizable
rithmic complexity would not translate into realiz- for spatial modeling in 2 or 3 dimensions and in
able performance benefits. In addition, the increase cases in which the variability is dominantly linear
in model parameters would render the models and (e.g., as modeled through a variogram or correlation
results difficult to interpret, and extrapolations likely matrix and because sampling through volumetric
unreliable. The desire to extrapolate is also a averaging over space introduces linearity through
stronger consideration in geostatistics than in ML, the central limit theorem, see e.g., Hsieh, 2002). The
especially since spatial correlation permits some kNN algorithm uses a variably weighted and aver-
extent of knowledge outside of sampled areas and aged value of the nearest ‘‘k’’ neighbors (a model
many ML algorithms extrapolate poorly or not at all hyperparameter) to determine the value of an un-
(e.g., tree-based methods). These characteristics known data point. This is very different to inverse
imply that a relatively small variety of known and distance modeling because distance is not a fixed
common ML algorithms may be useful for block parameter for kNN. In this paper, we focus on the
modeling without algorithmic modifications. systematic evaluation of the kNN algorithm for
Based on intuition from geostatistics, there are block modeling as a proof-of-concept. The hyper-
two classes of common ML algorithms that are parameter grid for the kNN algorithm was k = (2 to
theoretically appropriate for this task: neighbor- 30 in intervals of 1) and the weighting method was
based (e.g., k-nearest neighbors or kNN; see Fix & based on the inverse neighbor distance. During
Hodges, 1951; Cover & Hart, 1967; Witten & Frank, model selection, 20% of the sampled data was re-
Spatial Interpolation Using Machine Learning: from Patterns 135

served for testing and the remainder was used for (Krige, 1997); (2) kriging efficiency (KE; Krige 1997;
model training. During hyperparameter tuning, 4- Deutsch & Deutsch, 2012), which is a metric of the
fold cross-validation combined with the coefficient effectiveness of kriging estimates; (3) kriging slope of
of determination (CoD or R2) metric was used to regression (e.g., Deutsch et al., 2014), which is
select the best model. The implementation was in a metric used to estimate the local slope of a variable
Python using the Scikit-Learn library (Buitinck of interest in spatial interpolation. Based on a number
et al., 2013). of case studies, there is a correlation between the KE
and the kriging slope of regression (Krige, 1997).
However, these metrics are strictly within the disci-
Geostatistical Block Modeling pline of geostatistics and no direct counterpart of
them exist in ML. Hence, they cannot be used to cross-
Geostatistical block modeling is a very mature compare ML- and OK-based models. In addition,
task in spatial data modeling. Here, we provide a these metrics also cannot be used to compare models
simplified outline of the process for brevity. To with the synthetic ground truth. For this reason, we
perform geostatistical block modeling, we used a did not employ standard kriging metrics in this study.
standard geostatistical spatial estimation approach Instead, we made use of metrics from ML to compare
based on ordinary kriging (OK), which was imple- models against ground truth.
mented in GSLIB (Deutsch & Journel, 1992) and In this study, we employed block OK with block
Gestates (Pyrcz et al., 2021). OK is a method of parameters that matched with the synthetic data
spatial interpolation that uses a weighted linear generation, and also, notably, the use of a spherical
combination of sample values at nearby locations to model (Krige, 1997; Olea, 1999). This helped to
estimate the value at an unsampled location. The maximize the performance of block OK in a syn-
kriging process begins by defining a set of sample thetic study setting, because the data were synthet-
locations, known in geostatistics as the Ôneighbor- ically generated using a spherical variogram model;
hoodÕ or Ôsearch neighborhoodÕ, and the corre- choosing a spherical variogram model during block
sponding values of the variable at those locations modeling maximizes the accuracy of the resulting
(Krige, 1951). It is worth noting that the vernacular block models. However, in general, no ground truth
from various geoscientific disciplines overlap and is available and the choice of the variogram model is
sometimes conflict with those of artificial intelli- heuristic. The variogram model is a key subjective
gence, ML and data science. Neighborhood in the choice in a typical geostatistical workflow for inter-
sense of the ML algorithms does not refer to a polation because its structure cannot generally be
spatially constrained neighborhood, in general, and deduced unambiguously from data alone and no
its context depends on the definitions of the features. general knowledge is universally reliable. Maximiz-
The kriging estimate at an unsampled location is ing the performance of block OK is important to
then calculated as a weighted average of the values fully appreciate performance contrasts between the
at the sample locations within the neighborhood traditional OK approach and our proposed ML-
(Goovaerts, 1997). The weights are calculated such based approach. The tuning of model hyperparam-
that they minimize the estimation error, which is eters is automated and includes optimization of the
measured by the variance of the residuals (Krige & number of neighbors to use and fitting the variogram
Magri, 1982). In the case of OK, the sum of the following Zhang et al. (2023). In particular, the
weights for the individual samples is constrained to number of neighbors was selected using the elbow
unity and a Lagrange multiplier method is used to method combined with the mean absolute error
find a solution. Thus, the known mean of samples is metric, and the variogram was fitted using an auto-
not required in OK. There are also underlying mated least-squares method with a linear bias to-
assumptions, such as the stationarity of domains and ward shorter distances to favor better fitting of the
some extent of subjectivity in the form of modeling pre-range portion of the spherical model.
choices. These can greatly impact the results and the
interpretation of the kriging estimates.
There are a number of standard metrics in krig- Performance Assessment
ing to profile resulting model performance. These
include: (1) kriging variance, which is a measurement To assess the performance of both ML-derived
of the precision of an estimate at a given location and geostatistical block models by comparing the
136 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

resulting block models against the synthetic ground chain, metallurgical responses and proxy variables
truth, we used a variety of metrics that included the need to be included in the block model, which is
coefficient of determination (CoD or R2), the mean essential for optimizing mine planning and down-
absolute percentage error (MAPE, although it is stream mineral processing. This enriched block
expressed as a fraction by default, which was the model, known as the Ôgeometallurgical block modelÕ
case in this study as well; see https://scikit-learn.org/ (GMBM), is based on the transfer of simulated at-
stable/modules/generated/sklearn.metrics.mean_abs tributes (Deutsch et al., 2016; Garrido et al., 2017,
olute_percentage_error.html), and the median 2020). As there is now a general awareness of the
absolute error (MedAE; see https://scikit-learn.org/ lengthy exploration-to-extraction timelines and that
stable/modules/generated/sklearn.metrics.median_a a majority of the energy cost lies in beneficiation for
bsolute_error.html). In addition, we examined key many commodities and deposit types, there is a
statistical moments that included the mean and corresponding desire to drastically integrate the
standard deviation (STD) of all models. Lastly, to mineral value chain and enhance its agility. An early
assess the degree of smoothing in block models, we availability of GMBM is likely to be key to mineral
used a linear version of the dynamic range metric, value chain integration and enhanced agility because
which is defined as the difference between the it permits downstream metallurgical process designs
highest and the lowest block values. The dynamic early and, therefore, design of industrial processes
range metric is usually used in sensor engineering, and construction of facilities could be implemented
acoustic and image processing to measure sensor early. However, access to large mining exploration
and digital processing capability (https://www.electr and geometallurgical databases for method devel-
opedia.org/iev/iev.nsf/display?openform&ievref=72 opment and academic purposes is difficult due to
3-03-11). In our case, we can treat the resulting block confidentiality restrictions or budget limitations. To
models as images. Hence, the dynamic range metric address this issue and for the purpose of an academic
captured the maximum contrast of block models, study, we used a synthetic geometallurgical database
which is a useful metric of the level of smoothing of a typical porphyry Cu deposit (Garrido et al.,
because smoothing reduces intra-model contrast. To 2020) for a prototype deployment of our method.
ensure that the results were robust, we performed 50 This database contains geological and spatial coor-
runs for each combination of parameters and aver- dinates, mineralogical data, chemical assays and
aged the results, with 50 synthetic datasets being geometallurgical responses, such as rougher recov-
generated by varying the random number gen- ery and Bond work index (BWI). The BWI is a
eratorÕs seed. During each run, a single workflow measure of the resistance of the ore to grinding
was used, which created reproducible synthetic data using a laboratory ball mill. It is a widely accepted
conditions, including the random number seed and measure of ore grindability, indicating the energy
sampling pattern. The samples were then used to required to grind the ore to a specified size. The
perform block modeling and performance assess- BWI is expressed in units of kilowatt-hours per ton
ment. Metric results were thereafter averaged over of crushed material (kWh/tc). The mineralogical
the 50 runs to produce a statistically robust outcome. data include the proportions of chalcopyrite, bor-
nite, grouped clay minerals and pyrite in volume
percentage (vol.%), whereas chemical data include
Method Deployment Case and Data Description Cu grade in weight percent (wt.%). Boreholes at
various angles and lengths were extracted from the
Geometallurgy is the integration of geological, simulated deposit and processed for further model-
mineralogical, financial and metallurgical data ing, such as the spatial distribution of minerals,
(especially extractive metallurgy) in (usually) 3D elements, and geometallurgical responses. Figure 1
space to create a spatially-aware predictive model of shows the configuration of boreholes used for sub-
mineral processing. The benefits of geometallurgy in sequent modeling. It is important to note that many
mining engineering are significant, including geometallurgical variables are non-additive, such as
improvements in ore quality, mine planning, plant comminution indices, and combinations of compo-
performance, cost reduction and product quality sitional and non-compositional data (Deutsch et al.,
improvement (Jackson et al., 2011; Ortiz et al., 2015; 2016). Directly interpolating non-additive variables
Dominy et al., 2018; Garrido et al., 2019). To can result in misleading interpretations, although
incorporate these benefits into the mining value non-additive variables may not have the same im-
Spatial Interpolation Using Machine Learning: from Patterns 137

Figure 1. Graphical visualization of oriented boreholes with pyrite (vol.%) depicted (n = 153) with
variable azimuth and dip.

pact on all ML algorithms. However, for the els were visibly smoother in the interiors of high and
deployment portion of this study, the focus was not low concentration regions, as compared with the
to validate our method for non-additive variables in ground truth and the macroblocks (e.g., blue and red
general, but to provide a quantitative comparison zones in Figure 2). The qualitative differences be-
and examine the feasibility of our method in 3D came larger between block kriging and kNN results
using physically realistic data. at higher nugget effects (e.g., Fig. 4). Increasing le-
vels of anisotropy did not appear to exhibit a sys-
tematic qualitative effect (e.g., Fig. 5).
RESULTS Quantitatively, with increasing sample spacing,
there was an appreciable performance loss in both
Regular Sampling kNN and kriged models, although the CoD, MAPE
and MedAE metrics indicated that, on average, the
Qualitatively, under regular sampling condi- kNN models were closer to the synthetic data
tions, block OK produces models that are generally (Fig. 6). The dynamic range was systematically
higher in contrast (Figs. 2, 3, 4, 5). However, there is higher in the kriged models, which was consistent
a visually discernible loss of fine-scale detail (e.g., with the departure of STD scores from those of the
Fig. 3) in the kriged models as compared with that of samples (Fig. 6). This observation was consistent
kNN models. The difference is noticeable in Fig- with the qualitative finding that kNN models exhibit
ure 2 (compare the hotspot in the lower left cor- less contrast compared to kriged models. In general,
ners). However, it was not obvious whether the with increasing nugget effect, there is a gradual loss
increased detail retrieval resulted in overall more in performance of both kriged and kNN models,
accurate models, as it may be accompanied by a although the losses in CoD and MedAE are more
proportionate increase in noise as well. An observ- extensive with the kNN models (Fig. 7). The dy-
able qualitative difference was that block OK mod- namic range and STD scores also degraded slightly
138 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 2. Ground truth (synthetic 2D block model) and results of the kNN (ML microblocks and macroblocks) and kriged
modeling under regular sampling conditions. Spacing between samples is 40 cells.

more rapidly with the kNN models as compared to block modeling method was less accurate at the
the kriged models (Fig. 7). There seemed to be no distribution mean and median but it was capable of
robust systematic patterns of either the kNN or retrieving more fine-scale detail, although it seemed
kriged models in the metric scores with changes in to suffer from more smoothing despite increased
anisotropy (Fig. 8). These findings indicate that un- levels of fine detail.
der regular sampling conditions, block OK yields
generally more accurate models but at the expense
of a loss of fine detail. For other applications, it Biased Random Sampling
would be possible to tune the kriging neighborhood
to emphasize local detail, although this is not gen- The results of the kNN and kriged modeling
erally automatable. In contrast, the kNN-based under biased random sampling conditions were lar-
Spatial Interpolation Using Machine Learning: from Patterns 139

Figure 3. Ground truth (synthetic 2D block model) and results of the kNN (ML microblocks and macroblocks) and kriged
modeling under regular sampling conditions. Spacing between samples is 24 cells.

gely similar to those of regular sampling (Figs. 9, 10, conditions. In particular, with changes in the number
11, 12, 13 compared with Figs. 2, 3, 4, 5). In some of random samples, the kNN models were able to
cases, with strong anisotropy, the kNN models retrieve a greater dynamic range and more accurate
visually outperformed the kriged models qualita- STD values, at the expense of lower CoD, MAPE
tively in the sense that they reproduced variability and MedAE scores (Fig. 14). This was generally true
better across all azimuths (e.g., Fig. 13 but also less as well for the nugget effect sweep, although at very
pronounced in Fig. 12). This effect was not as low nugget values (< 0.1), the dynamic range per-
obvious under regular sampling conditions. Quanti- formance of the kNN and kriged models were re-
tative results were interesting in the sense that they versed in ranking but similar (Fig. 15). Although
were the reverse of those under regular sampling qualitatively obvious differences were noticeable
140 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 4. Ground truth (synthetic 2D block model) and results of the kNN (ML microblocks and macroblocks) and kriged
modeling under regular sampling conditions. Here, the nugget effect corresponds to 0.65.

with high anisotropy (e.g., Fig. 13), the metric scores Method Deployment
did not indicate a substantial or measurable differ-
ence (Fig. 16). These findings indicate that under The synthetic data used in this study had a clear
biased random sampling conditions, the kNN-based distinction between the different ore minerals, such
block modeling method is able to generally retrieve as bornite and chalcopyrite, enabling block model-
more detail but at the expense of also more noise. ing of each mineral volume proportion. The pro-
However, unlike the quantitative results under the posed microblocking approach demonstrated its
condition of regular sampling, the dynamic range capability to reproduce the data composition at the
and STD scores of the kNN models were systemat- sampled and block location during the method
ically better than those of the kriged models. development stage. During the visualization of the
Spatial Interpolation Using Machine Learning: from Patterns 141

Figure 5. Ground truth (synthetic 2D block model) and results of the kNN (ML microblocks and macroblocks) and kriged
modeling under regular sampling conditions. Anisotropy corresponds to 0.1.

3D models, a Ômoiré patternÕ or Ômoiré fringesÕ, that allowed higher material concentration to have
which are large-scale interference patterns that can larger cubes and lower concentration material to
occur when a partially opaque ruled pattern with have smaller and more transparent cubes. Figure 17
transparent gaps is overlaid on another similar pat- illustrates the predicted synthetic chalcopyrite pro-
tern, were observed. This was not related to the portion (vol.%) and its distribution in the studied
modeling methods prior to visualization and it was porphyry Cu deposit. The results of the prediction
strictly related to visualization of repeated or indicate that chalcopyrite is heavily concentrated
stacking patterns of 3D models when viewed in 2D toward the center of the deposit and on the surficial
with transparency. To mitigate this effect, we em- frontier of the deposit. The average chalcopyrite
ployed a scale-factor-based visualization technique content of the samples was 0.736 vol.% and it was
142 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 6. Parameter sweep of spacing (in cells) of regular sampling and impacts on (a) dynamic range, (b) CoD, (c) mean,
(d), MAPE, (e) standard deviation, and (f) MedAE.

comparable to the ML estimate of 0.739 vol.% tribution and concentration of bornite in the por-
(Table 2). These findings are consistent with the phyry Cu deposit. Our analysis revealed that bornite
simulated boreholes and ore deposit. The faint col- exhibits similar spatial patterns to chalcopyrite, al-
ors at the margins of the exploration area clearly beit with different grade distribution regularities
illustrate the extent of mineralization (Fig. 17). (Fig. 18). The average bornite content of sampled
The microblock modeling approach employed boreholes was 0.124 vol.% and the value was
in this study also allowed us to investigate the dis- reproduced to 0.124 vol.% in the ML estimates, with
Spatial Interpolation Using Machine Learning: from Patterns 143

Figure 7. Parameter sweep of nugget effect of regular sampling and impacts on (a) dynamic range, (b) CoD, (c) mean, (d),
MAPE, (e) standard deviation, and (f) MedAE.

the highest content clusters co-located in areas (Tosdal and Richards, 2001; Richards, 2003). These
where the content of chalcopyrite is high (Table 2). processes involve a phase separation in the fluid
These findings are consistent with the known dis- following a pressure-release caused by hydraulic
tribution of the ore minerals governed by changing fracturing of the country rock. During this process,
physico-chemical conditions during mineralization the Cu-rich fluids are generated by fractional crys-
and can be explained by a sequential paragenetic tallization of the magma, and subsequent migration
sequence. In porphyry Cu deposits, chalcopyrite and through fractures and faults in the surrounding rocks
bornite form by magmatic-hydrothermal processes deposit chalcopyrite and bornite. The formation of
144 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 8. Parameter sweep of anisotropy of regular sampling and impacts on (a) dynamic range, (b) CoD, (c) mean, (d),
MAPE, (e) standard deviation, and (f) MedAE.

chalcopyrite typically occurs earlier in the parage- deposit, 3D modeling of the Cu grade (wt.%)
netic sequence, followed by the formation of bor- showed that the host rock is enriched in Cu. The
nite. As such, bornite content in a given area can be average Cu content in the sampled boreholes was
linked to chalcopyrite content in that same area 0.369 wt.% and it was comparable to the predicted
(Fig. 18). estimate of 0.371 wt.% (Table 2). The distribution of
The Cu grade in a porphyry Cu deposit is Cu appears to follow a typical porphyry-style dis-
strongly related to the presence and concentration of tribution pattern (Fig. 19). Porphyry Cu deposits
Cu-bearing minerals. In the simulated porphyry Cu typically contain Cu minerals such as chalcocite,
Spatial Interpolation Using Machine Learning: from Patterns 145

Figure 9. Ground truth (synthetic 2D block model) and results of the kNN (ML microblocks and macroblocks) and kriged
modeling under biased random sampling conditions. The number of samples equals 625.

bornite, chalcopyrite and tennantite. Among these amenability to separation using froth flotation. The
minerals, chalcopyrite and bornite are considered deportment of Cu to different ore minerals is an
the primary Cu-bearing minerals, and their concen- important consideration in metallurgical processing.
tration is critical for Cu recovery during metallur- For example, Cu minerals such as chalcopyrite are
gical processing. These minerals typically occur as often associated with pyrite, which can create
disseminations or veinlets within the host rock. problems during processing due to the need for
Understanding Cu deportment to different ore selective flotation. Chalcopyrite is also more difficult
minerals is essential in processing of porphyry Cu to recover due to its low natural floatability, high
deposits. Bornite is considered a significant con- reactivity to oxidizing agents, and association with
tributor to Cu recovery due to its high Cu grade and pyrite and other gangue minerals. Additionally,
146 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 10. Ground truth (synthetic 2D block model) and results of the kNN (ML microblocks and macroblocks) and kriged
modeling under biased random sampling conditions. Number of samples equals 1736.

other minerals such as molybdenite can also impact clay minerals are present, accounting for up to
the metallurgical behavior of Cu in the ore. There- 15 vol.% of the total mineralogy of the sampled
fore, understanding the distribution of minerals and drillcores. The clay content in sampled boreholes
deportment of Cu is essential for optimizing metal- averaged 3.797 vol.% – a value that is similar to the
lurgical processing and maximizing Cu recovery. synthetic orebody composition and comparable to
Clay group minerals are known to adversely our ML estimated clay content of 3.849 vol.% in this
affect mineral processing, particularly flotation, due study. These minerals are variably distributed, and
to their ability to adsorb reagents and reduce min- some are in contact with Cu-bearing minerals, which
eral recovery. In the simulated porphyry Cu deposit, could further complicate mineral processing. Pre-
Spatial Interpolation Using Machine Learning: from Patterns 147

Figure 11. Ground truth (synthetic 2D block model) and results of the kNN (ML microblocks and macroblocks) and kriged
modeling under biased random sampling conditions. Here, the nugget corresponds to 0.65.

dictive geometallurgy of clay minerals can help can also help the design of targeted testing programs
understand their distribution and possible effects on to optimize flotation recovery in the presence of clay
mineral processing (Fig. 20). By incorporating min- minerals. Mitigating clay mineral factors regarding
eralogical data into a block model, process engineers flotation includes the use of dispersants and
can identify areas with higher clay contents and depressants to minimize clay interactions (i.e., this
modify their flotation reagent schemes accordingly. depends strongly on the clay mineralogy) with re-
In practical geometallurgical studies, it is more agents and minerals, respectively. Also, modifica-
important to model individual clay mineral contents tions to the flotation circuit design, such as
(i.e., how much kaolinite, sericite, montmorillonite, increasing the residence time or improving the froth
etc.) as opposed to total clay content. This approach stability, can help to mitigate the adverse effects of
148 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 12. Ground truth (synthetic 2D block model) and results of the kNN (ML microblocks and macroblocks) and kriged
modeling under biased random sampling conditions. Here, the nugget corresponds to 0.1.

clay minerals on flotation recovery. The use of pre- sampled boreholes was 13.264 kWh/tc, whereas the
dictive geometallurgy to identify areas with high clay ML modeled value was 13.432 kWh/tc and for the
content can facilitate the design of targeted testing synthetic deposit, 13.210 kWh/tc (Table 2). The BWI
programs to optimize flotation recovery in the is affected by the mineralogical composition of the
presence of these minerals (Fig. 20). For practical ore, as different minerals have different strengths,
geometallurgical studies, a much higher resolution and therefore, responses to grinding (Fig. 21). For
to capture the textural relationship between clays example, the presence of hard minerals such as
and sulfides will be required. quartz or pyrite can increase the BWI. In contrast,
The BWI for the simulated porphyry Cu deposit the presence of softer minerals such as clay minerals
ranged from 5 to 25 kWh/tc. The average BWI of the can decrease it. The BWI has important implications
Spatial Interpolation Using Machine Learning: from Patterns 149

Figure 13. Ground truth (synthetic 2D block model) and results of the kNN (ML microblocks and macroblocks) and kriged
modeling under biased random sampling conditions. Anisotropy corresponds to 0.1.

for the ore processing, as it affects the amount of geometallurgical block model, it is possible to opti-
energy required for grinding, which is a major con- mize the grinding circuit and reduce energy costs.
tributor to the cost of ore processing. It also affects Additionally, the BWI can be used to identify areas
the durability of grinding equipment, as higher BWI of the deposit with higher or lower energy require-
values can result in greater wear on the grinding ments, allowing for more efficient mine planning and
media and equipment. Predictive geometallurgy of improved processing.
the BWI is an important tool for understanding the The rougher recovery of a porphyry Cu deposit
properties of the ore and predicting expected energy is a critical metallurgical response variable that can
costs. Through incorporating the BWI into a significantly impact the economics of mining and
150 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 14. Parameter sweep of number of samples of biased sampling and impacts on (a) dynamic range, (b) CoD, (c) mean,
(d), MAPE, (e) standard deviation, and (f) MedAE.

processing operations. In the present study, the mineralogy of the deposit. Average rougher recov-
rougher recovery results of the simulated porphyry ery of sampled boreholes was 83.384% while the
Cu deposit indicated that the ore was amenable to actual recovery was 83.361%. These values are clo-
metallurgical treatment with rougher recovery sely approximated in the estimate and average at
ranging from 75 to 95%. The range of recovery 84.538% (Table 2). However, some areas of the
values was considered to be quite good for this type deposit have lower recovery rates, i.e., less than
of deposit, which is primarily due to the favorable 70%, which is attributed to the complex mineralogy
Spatial Interpolation Using Machine Learning: from Patterns 151

Figure 15. Parameter sweep of nugget effect of samples of biased sampling and impacts on (a) dynamic range, (b) CoD, (c)
mean, (d), MAPE, (e) standard deviation, and (f) MedAE.

and alteration of the host rocks by clay minerals composition and distribution of clay minerals in the
(Fig. 22). This indicates that the presence of clay ore, which can help in designing appropriate pro-
minerals can negatively impact the recovery of cessing strategies to mitigate the adverse effects of
valuable minerals during metallurgical processing. such minerals. In addition, the rougher recovery
This underscores the importance of predictive response variable can also be used to predict ore
geometallurgy in understanding the mineralogical properties and expected recoveries, which can, in
152 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 16. Parameter sweep of anisotropy of biased sampling and impacts on (a) dynamic range, (b) CoD, (c) mean, (d),
MAPE, (e) standard deviation, and (f) MedAE.

turn, impact the selection of processing equipment, DISCUSSION

as well as the energy and cost requirements of pro-
cessing. Therefore, incorporating a rougher recovery Performance Summary and Implications
response variable into the block model can help
optimize the entire mining value chain from explo- Block modeling is a common and important
ration to processing. task in a variety of geoscientific, mining and related
Spatial Interpolation Using Machine Learning: from Patterns 153

Figure 17. Predicted 3D block model showing the distribution of chalcopyrite (vol.%).

Table 2. Statistical quantities of the samples, predicted and actual 3D block models. The units for all minerals are in vol.% and for Cu, in
wt.%

Materials Mean STD Min 25% 50% 75% Max

Chalcopyrite (samples) 0.736 0.541 0 0.411 0.648 0.917 7.430

Chalcopyrite (actual block model) 0.728 0.528 0 0.414 0.645 0.906 7.583
Chalcopyrite (predicted block model) 0.739 0.461 0 0.478 0.674 0.896 6.016
Bornite (samples) 0.124 0.151 0 0.023 0.084 0.166 2.292
Bornite (actual block model) 0.122 0.149 0 0.023 0.083 0.163 3.445
Bornite (predicted block model) 0.124 0.111 0 0.053 0.100 0.160 1.548
Cu (samples) 0.369 0.240 0 0.225 0.341 0.470 3.698
Cu (Actual block model) 0.364 0.236 0 0.225 0.338 0.462 4.060
Cu (predicted block model) 0.371 0.198 0 0.262 0.357 0.457 2.777
Clay minerals (samples) 3.797 2.897 0 1.827 3.087 5.078 40.670
Clay minerals (actual block model) 3.797 2.888 0 1.836 3.092 5.072 40.670
Clay minerals (predicted block model) 3.849 2.155 0 2.371 3.480 4.947 30.232
BWI (samples) 13.264 3.734 0 12.600 12.921 13.335 28.524
BWI (actual block model) 13.210 3.669 0 12.602 12.920 13.323 28.567
BWI (predicted block model) 13.432 2.888 0 12.650 12.914 13.324 27.896
Rougher recovery (samples) 83.384 16.947 0 84.197 87.358 89.399 94.368
Rougher recovery (actual block model) 83.361 17.008 0 84.178 87.343 89.383 94.857
Rougher recovery (predicted block model) 84.538 11.139 0 84.735 86.915 88.524 93.356

engineering activities. Block kriging is the dominant parametric model of the spatial variability structure
solution of the change-of-support problem by pop- to produce a weighted average. In comparison,
ularity in mapping and resource estimation. It uses a current varieties of ML algorithms and, in particular,
154 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 18. Predicted 3D block model showing the distribution of bornite (vol.%).

the kNN algorithm do not explicitly model spatial performance differences and implications, it is nec-
variability. Empirically, it is clear that under the essary to venture beyond empirical results and
conditions that were used for synthetic data gener- examine the algorithm architecture. As the present
ation and sampling, the performance of the kNN applicative overlap between ML and geostatistics is
algorithm in our microblock-to-macroblock method still small, we provide a discussion from two per-
varies depending on the specifics of the sampling spectives. The first perspective examines the algo-
practice and the metrics considered. Under regular rithms in context of geostatistics, while the second
sampling conditions, the kNN algorithm was more perspective examines the algorithms in the context
accurate at the mean to median level, as evidenced of ML.
by the generally higher CoD, lower MAPE and The architecture of the kNN algorithm can be
MedAE scores (Figs. 6, 7, 8). However, the level of effectively examined through the perspective of
smoothing was generally lower for block kriging in geostatistics to understand the observed perfor-
comparison with the kNN algorithm (Figs. 6, 7, 8). mance differences. The kNN algorithm uses the
This trend is reversed for results under biased sam- average of multiple neighbors, which in the case of
pling conditions (Figs. 14, 15, 16). This implies that prediction of each block within the microblock grid,
under realistic conditions that are somewhere be- means that the closest k-members of each micro-
tween completely regular and completely random block is used to derive an average. The idea of a
sampling, there is no strong rationale to pick one neighborhood in ML is a generic one and only under
method over the other for the purpose of block the special circumstance that spatial coordinates are
modeling. Within the explored conditions in this the sole features, is the concept of neighborhood
study, the performance differences as measured by identical to that in geostatistics. This implies that the
the CoD, MAPE and MedAE metrics are generally support for each microblock is a fixed number of
small (typically on the order of 10-2 or a few percent neighbors, which in the case of regular sampling, is a
absolute difference). To further understand the spatial neighborhood of fixed radius. In the case of
Spatial Interpolation Using Machine Learning: from Patterns 155

Figure 19. Predicted 3D block model showing the distribution of Cu (wt.%).

biased random sampling, this changes completely case in inverse distance weighting. In other words,
because the nearest neighbors do not occur at fixed setting the hyperparameter ‘‘k’’ is equivalent to
distance intervals, and, therefore, the support is not optimizing the number of neighbors to use in krig-
a spatially consistent concept, and no fixed spatial ing, although in the latter, the model-tuning process
neighborhood is applicable. In our experience, the is not necessarily cross-validation based. Fixing the
biased random sampling is more representative of number of neighbors through cross-validation im-
actual data, and hence, the concept of a fixed plies that relative to inverse distance modeling, kNN
‘‘search’’ radius in the case of block kriging does not is more resistant to over-smoothing (because of ex-
generally have an equivalent in most ML algorithms. plicit controls on model bias and variance). In fact,
The way in which averaging of neighboring points is the retrieval of fine-scale detail, particularly under
employed in kNN can technically be either un- more challenging conditions of biased random
weighted or weighted. In this study, we used dis- sampling, higher nugget effect and anisotropy (e.g.,
tance-weighting, which implies that contributions of Figs. 11 and 13) is indicative of a reduction of re-
points are inversely related to their distance. This is liance on distal supports.
mechanistically similar to inverse distance weight- The proposed microblocking approach using
ing, which is a type of spatial modeling algorithm ML has demonstrated its effectiveness in reproduc-
that is simpler than kriging. However, a key differ- ing ore distribution patterns and regularities in the
ence between kNN and inverse distance weighting is simulated porphyry Cu deposit (Figs. 18 and 22).
that kNN uses explicit control on the notion of the This approach was then used to create 3D predictive
size of neighborhood through a single hyperparam- geometallurgical models that incorporate mineral-
eter – the number of neighbors. This means that the ogy, chemistry and metallurgical response variables,
delineation of the boundary of neighborhood is which can be used to improve mining, mine plan-
sharp in the case of kNN, which his generally not the ning, grade control and mineral processing and
156 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 20. Predicted 3D block model showing the distribution of clay minerals (vol.%).

metallurgical extraction (e.g., Fig. 22). Using the ing circuits that maximize recovery while minimizing
microblocking approach, it was possible to accu- energy consumption and equipment wear.
rately reproduce the concentration of the ore at From the ML perspective, the main task of a
sampled and block locations of the synthetic data ML model is to capture as much information as
during the deployment testing stages (Table 2). The possible that is contained within the given data.
predictive geometallurgical models generated using Block modeling is a highly specific application of
the microblocking approach provided valuable in- ML in this sense, because in this case, the spatial
sights into the distribution of Cu, bornite, chal- information is captured by a ML model and repli-
copyrite and other minerals in the selected porphyry cated onto a fine microblock grid. This is similar to
Cu deposit. These models also allowed for the the classification application of domain boundary
identification of areas with complex mineralogy and delineation using ML (Zhang et al., 2023). The
alteration by clay minerals, which could result in performance of a ML model depends on its ability to
lower recovery rates during metallurgical treatment. model the relationships between the features and
This information can be used to optimize mining, the data label contained within the data. In this
mine planning and grade control strategies to max- sense, an appropriate choice of a ML algorithm
imize the recovery of valuable minerals while mini- should yield a minimum of bias and variance, such
mizing costs. Moreover, the predictive that the performance of the resulting models is
geometallurgical models created using the maximized. For spatial learning tasks, the concept of
microblocking approach can be used to optimize local continuity of the spatial variability, or spatial
mineral processing and metallurgical extraction correlation, is not explicitly addressed in known ML
processes. By incorporating information about the algorithms. In the case of the kNN algorithm, spatial
mineralogy, chemistry and metallurgical response of learning is possible because the algorithm uses a
the deposit, it is possible to design optimal process- notion of a local and weighted neighborhood com-
bined with the use of distance metrics and spatial
Spatial Interpolation Using Machine Learning: from Patterns 157

Figure 21. Predicted 3D block model of bond work index (BWI; kwh/tc).

dimensions as features. Perfect replication of data is implies that methods that do not assume global data
theoretically possible to within the constraints properties or require geodomaining could be gen-
placed by sampling (e.g., resolution). As the data erally more applicable, in that they require less
noise increases, for example, through increase of the considerations and treatise of the data prior to block
nugget effect, the performance of models decreases modeling. Although this remains to be verified
(e.g., Figs. 7 and 15). This is best understood from empirically. This is the case for the kNN algorithm
the error decomposition into model bias, variance at least in the realm of data science because we are
and irreducible error. Increasing the nugget effect unaware of any requirement to segment training
increases irreducible error. data prior to the use of kNN-based classification or
Despite the advantages of block kriging, regression, although in specific cases (e.g., where
including a very mature workflow and many appli- there are highly heterogeneous clusters and predic-
cation examples, there are some undesirable aspects tion occurs within clusters), this is likely to be ben-
of block kriging. One of the main challenges is its eficial. Additionally, fitting of variograms and
complicated and long sequence of procedures trimming of data tend to require expert judgment
(Cressie, 1990). Furthermore, both ML-based and and substantial disciplinary experience. It is unclear
block kriging methods still require large amounts of what the net uncertainty across this sequence of
data to produce accurate and reliable results, which procedures is because a systematic study has not
can be a challenge in areas with sparse or limited been conducted. However, block models must al-
data (Deutsch et al., 2014). Additionally, block ways be used with grade control and metal recon-
kriging of various forms assumes some spatial sta- ciliation, which are feedback mechanisms to address
tistical properties in the data (either global or at model uncertainty (Chiles & Delfiner, 2012). The
least satisfiable within geodomains, see Zhang et al., impacts of this type of uncertainty on mapping (e.g.,
2023), including that the geological variable being for exploration, including prospectivity mapping) is
modeled is stationary. This assumption may not al- wholly unknown, as reconciliation is not possible at
ways hold true in practice, particularly in heteroge- this scale and for this type of activity. Consequently,
neous geological environments. Theoretically, this the ability to minimize the length and complexity of
158 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Figure 22. Predicted 3D block model of rougher recovery (%).

the sequence of procedures toward usable block ing-based workflow, which is consistent with their
models will likely increase model reproducibility training. This benefit is likely to become more
and objectivity. In our case, the microblock-to- prominent with the rise of geodata science and
macroblock approach using the kNN algorithm higher-than-traditional velocity (ungridded)
makes use of a minimum of 2 key hyperparameters – geospatial data.
the size of microblocks and the number of neigh-
bors. The microblock size parameter is not as
impactful on the outcome, as it can be set as small as CONCLUSIONS
computationally practical (e.g., as large grids con-
sume more memory) and, in general, decreasing the Directly producing block models for a variety of
parameter value results in increasingly diminishing purposes that include mapping and mineral resource
returns. The tuning of the number of neighbors in estimation is challenging with ML algorithms be-
kNN is straightforward and fully data-driven and cause this task involves a change of support. ML
automated using the ML workflow, which means no algorithms implicitly assume that the prediction is
subjectivity is involved beyond data pre-processing made at the same size of support as the input and
(which was not employed in our study). This means training data, and hence, unlike block kriging,
that in comparison with block kriging, our method is changing support requires additional consideration.
essentially equally applicable to within the extent of In this paper, we purposefully exploited the constant
data explored but is substantially simpler and support property of ML algorithms and, in particu-
essentially requires no manual contribution, aside lar, explored the use of the simple kNN algorithm
from coding a workflow. This is an intended benefit for the purpose of block modeling. We demon-
of our approach, which enables data scientists to strated that under the conditions of regular and
interpolate geospatial data using a machine learn- biased random sampling, the kNN and microblock-
Spatial Interpolation Using Machine Learning: from Patterns 159

to-macroblock method yields results that are quali- through Natural Resources Canada (Geological
tatively and quantitatively similar to those of block Survey of Canada) was provided by the Critical
kriging. In addition, detailed differences are Minerals Geoscience Data (CMGD) program.
nuanced in terms of qualitative and quantitative
performance. We expect that our approach will be DECLARATIONS
useful under a range of typical conditions, especially
when highly reproducible and objective outcomes Conflict of Interest The authors declare that they
are expected, because our method requires tuning of have no known competing financial interests or
a substantially reduced set of parameters as com- personal relationships that could have appeared to
pared with the geostatistical approach. Our ap- influence the work reported in this paper.
proach recognizes that the idea of punctual support
is a mathematical approximation, which serves OPEN ACCESS
geostatistics well but is inconsequential to detri- This article is licensed under a Creative
mental to ML in spatial tasks. Reconsidering this Commons Attribution 4.0 International License,
assumption and, therefore, undoing the punctual which permits use, sharing, adaptation, distribution
approximation enables the construction of a micro- and reproduction in any medium or format, as long
block grid, such that the individual predictions can as you give appropriate credit to the original
be made at a scale that is much smaller than that of author(s) and the source, provide a link to the
the final macroblock grid size and in a manner that is Creative Commons licence, and indicate if changes
closer to the finite size of support as represented by were made. The images or other third party material
the data. The two key recognitions are: (1) that the in this article are included in the article’s Creative
solution requires significant multidisciplinary Commons licence, unless indicated otherwise in a
knowledge at the geoscientific level (as in outside of credit line to the material. If material is not included
the data science realm); and (2) modifying existing in the article’s Creative Commons licence and your
assumptions as necessary (notion of punctual sup- intended use is not permitted by statutory regulation
port) to fit with ML algorithms. Formulating a fully or exceeds the permitted use, you will need to obtain
automatable spatial interpolation method, with a permission directly from the copyright holder. To
workflow implemented in the ML framework would view a copy of this licence, visit http://creativecom
imply that our method should be highly accessible to mons.org/licenses/by/4.0/.
data scientists, who may wish to venture into geo-
sciences and perhaps, eventually become geodata
scientists.
REFERENCES

Abzalov, M. Z., & Humphreys, M. (2002). Resource estimation of

ACKNOWLEDGMENTS structurally complex and discontinuous mineralization using
non-linear geostatistics: case study of a mesothermal gold
deposit in northern Canada. Exploration and Mining Geol-
The authors would like to thank the reviewers ogy, 11(1–4), 19–29.
for their comments which have greatly improved this Annels, A. E. (1991). Mineral deposit evaluation, a practical ap-
manuscript. The authors also would like to thank proach. Chapman and Hall.
Armstrong, M., & Champigny, N. (1989). A study on kriging small
E.J.M. Carranza for his generous time and editorial blocks. CIM Bulletin, 82, 128–133.
handling. Breiman, L. (1996a). Bagging predictors. Machine Learning, 24,
123–140.
Breiman, L. (1996b). Stacked regressions. Machine Learning, 24,
49–64.
FUNDING Buitinck, L., Louppe, G., Blondel, M., Pedregosa, F., Mueller, A.,
Grisel, O., Niculae, V., Prettenhofer, P., Gramfort, A.,
Grobler, J., Layton, R., Vanderplas, J., Joly, A., Holt, B., &
Open access funding provided by University of Varoquaux, G. (2013). API design for machine learning
the Witwatersrand. Funding for this research software: experiences from the scikit-learn project. Retrieved
through the Wits Mining Institute was provided by May 13, 2023, from: http://arxiv.org/abs/1309.0238.
Carranza, E. J. M. (2009). Controls on mineral deposit occurrence
the DSI-National Research Foundation (NRF) inferred from analysis of their spatial pattern and spatial
Thuthuka Grant (Grant No: 121973) and DSI-NRF association with geological features. Ore Geology Reviews,
CIMERA. Additional funding for this research 35(3–4), 383–400.
160 Nwaila, Zhang, Bourdeau, Frimmel, and Ghorbani

Carvalho, D., & Deutsch, C. V. (2017). An overview of multiple deposits. In Procemin GEOMET 2019, 15th International
indicator kriging. Geostatistics Lessons. Retrieved May 13, Mineral Processing Conference, 6th International Seminar on
2023, from: http://geostatisticslessons.com/lessons/mikovervie Geometallurgy, Santiago, Chile.
w. Garrido, M., Sepúlveda, E., Ortiz, J., & Townley, B. (2020).
Cevik, I. S., Leuangthong, O., Cate, A., & Ortiz, J. M. (2021). On Simulation of synthetic exploration and geometallurgical
the use of machine learning for mineral resource classifica- database of porphyry copper deposits for educational pur-
tion. Mining, Metallurgy & Exploration, 38, 2055–2073. poses. Natural Resources Research, 29, 3527–3545.
Chiles, J., & Delfiner, P. (2012). Geostatistics: Modelling spatial Gelfand, A. E., Zhu, L., & Carlin, B. P. (2001). On the change of
uncertainty (2nd ed.). Wiley. support problem for spatio-temporal data. Biostatistics, 2(1),
Cover, T., & Hart, P. (1967). Nearest neighbor pattern classifi- 31–45.
cation. IEEE Transactions on Information Theory, 13, 21–27. Goertzel, B. (2006). The hidden pattern: A patternist philosophy of
Cressie, N. (1990). The origins of kriging. Mathematical Geology, mind. BrownWalker Press.
22, 239–252. Good, I. J. (1983). The philosophy of exploratory data analysis.
Crochiere, R. E., & Rabiner, L. R. (1983). Multirate digital signal Philosophy of Science, 50(2), 283–295.
processing (vol. 18). Prentice-Hall. Goovaerts, P. (1997). Geostatistics for natural resources evalua-
David, M., (1976). The practice of kriging. In Advanced Geo- tion. Oxford University Press.
statistics in the Mining Industry: Proceedings of the NATO Gotway, C. A., & Young, L. J. (2002). Combining incompatible
Advanced Study Institute held at the Istituto di Geologia spatial data. Journal of the American Statistical Association,
Applicata of the University of Rome, Italy, 13–25 October 97(458), 632–648.
1975 (pp. 31-48). Springer Netherlands. Groves, D. I., Vielreicher, R. M., Goldfarb, R. J., & Condie, K. C.
Dennett, D. C. (1991). Real patterns. The Journal of Philosophy, (2005). Controls on the heterogeneous distribution of mineral
88(1), 27–51. deposits through time. Geological Society, London, Special
Deutsch, J. L., & Deutsch, C. V. (2012). Kriging, stationary and Publications, 248, 71–101.
optimal estimation: measures and suggestions. CCG Annual Ho, T. K. (1995). Random decision forests. In Proceedings of the
Report 14, Paper 306. 3rd International Conference on Document Analysis and
Deutsch, C. V., & Journel, A. G. (1992). GSLIB: Geostatistical Recognition (Vol. 1, pp. 278–282). IEEE, Montréal, Canada.
software library and userÕs guide. Oxford University Press. https://doi.org/10.1109/ICDAR.1995.598994.
Deutsch, J., Palmer, K., Deutsch, C., Szymanski, J., & Etsell, T. Hsieh, W. W. (2002). The impact of time-averaging on the
(2016). Spatial modelling of geometallurgical properties: detectability of nonlinear empirical relations. Quarterly
Techniques and a case study. Natural Resources Research, 25, Journal of the Royal Meteorological Society: A journal of the
161–181. atmospheric sciences, applied meteorology and physical
Deutsch, J. L., Szymanski, J., & Deutsch, C. V. (2014). Checks and oceanography, 128(583), 1609–1622.
measures of performance for kriging estimates. The Journal Isaaks, E. (2005). The kriging oxymoron: a conditionally unbiased
of the Southern African Institute of Mining and Metallurgy, and accurate predictor (2nd Edition). In Leuangthong, O.,
114, 223–230. Deutsch, C.V. (Eds.), Geostatistics Banff 2004. Quantitative
Diaz-Gonzalez, F. A., Vuelvas, J., Correa, C. A., Vallejo, V. E., & Geology and Geostatistics, vol 14. Springer. https://doi.org/
Patino, D. (2022). Machine learning and remote sensing 10.1007/978-1-4020-3610-1_37.
techniques applied to estimate soil indicators–review. Eco- Isaaks, E. H., & Srivastava, R. M. (1989). An introduction to ap-
logical Indicators, 135, 108517. plied geostatistics. Oxford University Press.
Dominy, S., OÕConnor, L., Parbhakar-Fox, A., Glass, H., & Jackson, J., McFarlane, A., & Olson, K. (2011) Geometallurgy –
Purevgerel, S. (2018). Geometallurgy: A route to more re- back to the future: scoping and communicating geomet pro-
silient mine operations. Minerals, 8(12), 560. grams. In GeoMet 2011 – 1st AusIMM International
Dumakor-Dupey, N. K., & Arya, S. (2021). Machine learning – a Geometallurgy Conference 2011 (pp. 115–123). Australasian
review of applications in mineral resource estimation. En- Institute of Mining and Metallurgy.
ergies, 14(4), 4079. Kotsiantis, S. B. (2014). Integrating global and local application of
Fix, E., & Hodges, J. L. (1951). An important contribution to naive bayes classifier. International Arab Journal of Infor-
nonparametric discriminant analysis and density estimation. mation Technology, 11(3), 300–307.
International Statistical Review, 57(3), 233–238. Kotsiantis, S. B., Zaharakis, I., & Pintelas, P. (2007). Supervised
Freund, Y., & Schapire, R. E. (1997). A decision-theoretic gen- machine learning: A review of classification techniques.
eralization of on-line learning and an application to boosting. Emerging artificial intelligence applications in computer
Journal of Computer and System Sciences, 55(1), 119–139. engineering, 160(1), 3–24.
Galetakis, M., Vasileiou, A., Rogdaki, A., Deligiorgis, V., & Krige, D. G. (1951). A Statistical approach to some mine valua-
Raka, S. (2022). Estimation of mineral resources with ma- tions and allied problems at the Witwatersrand. M.Sc. thesis,
chine learning techniques. Materials Proceedings, 5(1), 122. University of Witwatersrand, Johannesburg, South Africa.
Garrido, M., Sepúlveda, E., & Navarro, F. (2017). Optimisation of Krige, D. G. (1997). A practical analysis of the effects of spatial
planning and scheduling of ore body with open pit extraction structure and of data available and accessed, on conditional
considering homogeneity in clays as geometallurgical vari- biases in Ordinary Kriging. In E. Y. Baafi & N. A. Schofield
ables. In Geomin Mine planning, 5th International Seminar on (Eds.), Geostatistics Wollongong 96. Quantitative Geology
Geology for the Mining Industry, 5th International Seminar and Geostatistics (Vol. 2, pp. 799–810). Kluwer Academic
on Mine Planning, Santiago, Chile. Press.
Garrido, M., Sepúlveda, E., Ortiz, J., Navarro, F., & Townley B. Krige, D. G., & Magri, E. J. (1982). Geostatistical case studies of
(2018). A methodology for the simulation of synthetic the advantages of lognormal-de Wijsian kriging with mean
geometallurgical block models of porphyry ore bodies. In for a base metal mine and a gold mine. Journal of the Inter-
Procemin GEOMET 2018, 14th International Mineral Pro- national Association for Mathematical Geology, 14, 547–555.
cessing Conference, 5th International Seminar on Geometal- Kuipers, T. A. (2001). Structures in science: heuristic patterns
lurgy, Santiago, Chile. based on cognitive structures. In An Advanced Textbook in
Garrido, M., Ortiz J., Sepúlveda, E., Farfan, L., & Townley, B. Neo-Classical Philosophy of Science (vol. 301). Springer.
(2019). An overview of good practices in the use of Lawley, C. J., McCafferty, A. E., Graham, G. E., Huston, D. L.,
geometallurgy to support mining reserves in copper sulfides Kelley, K. D., Czarnota, K., Paradis, S., Peter, J. M., Hay-
Spatial Interpolation Using Machine Learning: from Patterns 161

ward, N., Barlow, M., & Emsbo, P. (2022). Data–driven Richards, J. P. (2003). Tectono-magmatic precursors for porphyry
prospectivity modelling of sediment–hosted Zn–Pb mineral Cu-(Mo-Au) deposit formation. Economic Geology, 98(8),
systems and their critical raw materials. Ore Geology Re- 1515–1533.
views, 141, 104635. Sagi, O., & Rokach, L. (2018). Ensemble learning: A survey.
Lawley, C. J., Tschirhart, V., Smith, J. W., Pehrsson, S. J., Wiley Interdisciplinary Reviews: Data Mining and Knowledge
Schetselaar, E. M., Schaeffer, A. J., Houlé, M. G., & Discovery, 8(4), e1249.
Eglington, B. M. (2021). Prospectivity modelling of Canadian Samson, M. (2020). Mineral resource estimates with machine
magmatic Ni (±Cu±Co±PGE) sulphide mineral systems. learning and geostatistics. M.Sc. thesis, University of Alberta,
Ore Geology Reviews, 132, 103985. Canada. https://doi.org/10.7939/r3-xxxz-5z86.
Liu, Y., Carranza, E. J. M., & Xia, Q. (2022). Developments in Sarma, D. D. (2009). Kriging Variance and kriging procedure. In
quantitative assessment and modeling of mineral resource Geostatistics with applications in Earth Sciences (pp. 125-138).
potential: an overview. Natural Resources Research, 31, 1825– Springer. https://doi.org/10.1007/978-1-4020-9380-7_8.
1840. Srinivasan, K., & Fisher, D. (1995). Machine learning approaches
Matheron, G. (1967). Kriging or polynomial interpolation proce- to estimating software development effort. IEEE Transac-
dures. CIMM Transactions, 70(1), 240–244. tions on Software Engineering, 21(2), 126–137.
Maxwell, A. E., Warner, T. A., & Fang, F. (2018). Implementation Steiner, M. (2009). Empirical regularities in WittgensteinÕs phi-
of machine-learning classification in remote sensing: An ap- losophy of mathematics. Philosophia Mathematica, 17(1), 1–
plied review. International Journal of Remote Sensing, 39(9), 34.
2784–2817. Talebi, H., Mueller, U., Tolodana Delgado, R., & van den Boo-
Mery, N., & Marcotte, D. (2022). Quantifying mineral resources faart, K. G. (2019). Geostatistical simulation of geochemical
and their uncertainty using two existing machine learning compositions in the presence of multiple geological units:
methods. Mathematical Geosciences, 54, 363–387. Application to mineral resource evaluation. Mathematical
Nwaila, G. T., Zhang, S. E., Bourdeau, J. E., Ghorbani, Y., & Geosciences, 51, 129–153.
Carranza, E. J. M. (2022). Artificial intelligence-based Tosdal, R. M., & Richards, J. P. (2001). Magmatic and structural
anomaly detection of the Assen iron deposit in South Africa controls on the development of porphyry Cu±Mo±Au de-
using remote sensing data from the Landsat-8 Operational posits. In: JP Richards, RM Tosdal (Eds.), Structural controls
Land Imager. Artificial Intelligence in Geosciences, 3, 71–85. on ore genesis, Reviews in Economic Geology (vol. 14).
Olea, R. A. (1999). Geostatistics for engineers and earth scientists. Society of Economic Geologists.
Berlin: Springer. Uddin, M. N., & Hamiduzzaman, M. (2009). The philosophy of
Ortiz, J., Kracht, W., Townley, B., Lois, P., Cárdenas, E., Mir- science in social research. The Journal of International Social
anda, R., & Alvarez, M. (2015) Workflows in geometallur- Research, 2(6), 1–11.
gical prediction: challenges and outlook. In Proceedings of Vann, J., Jackson, S., & Bertoli, O. (2003). Quantitative kriging
the 17th Annual Conference of the International Association neighborhood analysis for the mining geologist – a descrip-
for Mathematical Geosciences IAMG 2015. tion of the method with worked case examples. In 5th
Ortiz, J. M., & Emery, X. (2006). Geostatistical estimation of International Mining Geology Conference (vol. 8, pp. 215-
mineral resources with soft geological boundaries a compar- 223). Bendigo, Australia. Melbourne: Australian Institute of
ative study. Journal of the South African Institute of Mining Mining & Metallurgy.
and Metallurgy, 106, 577–584. Veronesi, F., & Schillaci, C. (2019). Comparison between geo-
Parsa, M., Lentz, D. R., & Walker, J. A. (2023). Predictive statistical and machine learning models as predictors of
modeling of prospectivity for VHMS mineral deposits, topsoil organic carbon with a focus on local uncertainty
northeastern Bathurst mining camp, NB, Canada, using an estimation. Ecological Indicators, 101, 1032–1044.
ensemble regularization technique. Natural Resources Re- Washburn, D. K., & Crowe, D. W. (1988). Symmetries of culture:
search, 32, 19–36. Theory and practice of plane pattern analysis. University of
Pyrcz, M. J., Jo, H., Kupenko, A., Liu, W., Gigliotti, A. E., Washington Press.
Salomaki, T., & Santos, J. (2021). Gestates python package. Witten, I. H., & Frank, E. (2005). Data mining: practical machine
PyPI, Python Package Index. Retrieved May 13, 2023, from: learning tools and techniques (2nd ed.). Morgan Kaufman.
https://pypi.org/project/geostatspy/. Zhang, S. E., Nwaila, G. T., Bourdeau, J. E., Ghorbani, Y., &
Rasmussen, C. E., & Williams, C. K. (2006). Gaussian processes Carranza, E. J. M. (2023). Machine learning-based delin-
for machine learning (Vol. 1). Cambridge: MIT press. eation of geodomain boundaries: A proof-of-concept study
using data from the Witwatersrand Goldfields. Natural Re-
sources Research, 32, 879–900.