1. Introduction
Mineral resources are indispensable to the sustenance of modern civilization [
1,
2]. They play essential roles in socioeconomic development, industrial processes, manufacturing of modern technologies, and construction of modern transportation systems [
2,
3,
4,
5]. A mineral resource is commonly defined as “a concentration of naturally occurring material in or on the Earth’s crust in such form and amount that economic extraction of a commodity from the concentration is currently or potentially feasible” [
6,
7,
8]. Evaluation of these resources (mineral resource estimation) is a crucial and challenging task in every mineral exploration and mining project, irrespective of size, commodity, and deposit type [
9,
10]. Mineral resource estimation is performed to determine the quantity and quality of a mineral deposit and to establish confidence in its geological interpretation, and it requires careful and detailed consideration of high spatial variability and uncertainty associated with geological formations [
11]. Therefore, a reliable mineral resource estimate is critical to the success of every mining project.
Mineral resources are subdivided into inferred, indicated, and measured categories based on increasing geological confidence and knowledge, as illustrated in
Figure 1 [
8]. The estimation of a mineral resource is followed by a mineral reserve estimation, which is carried out to determine the tonnage and average grade of a mineral deposit that is economically and technologically feasible to mine [
12]. Mineral resource estimation underlies the generation of mineral reserve estimates. Mineral reserve estimate establishes the mineable portion of a resource and forms the foundation for economic anaslysis of a mineral deposit as well as the future potential of an operating mine. The accuracy of a reserve estimation is essential to the quality of the geological interpretation [
13,
14]. It is also vital to mine planning and design, including the utility of short-term and long-term mine plans. Moreover, the estimation accuracy is key to mining decisions, such as capital allocation, operating policy, depletion rate, and depreciation charges [
13,
15,
16]. Therefore, an accurate reserve estimation is critical to the feasibility, sustainability, and daily/future operations of a mining project. This phase of the estimation process is also referred to as ore reserve estimation and grade estimation [
17]. It should be noted that mineral resource or reserve estimates are not the only factors that determine the extractability of a mineral resource; there are other deciding factors, such as economic, environmental, climatic, and social restrictions [
6]. Mineral reserves are classified into probable and proved reserves.
Figure 1 shows a general classification of exploration results based on the levels of confidence in geological knowledge and technical and economic considerations about the deposit as established by the Australasian Code for Reporting of Identified Mineral Resources and Ore Reserves (The JORC Code) [
17].
There are two concepts underlying reserve estimation: the concept of extension, where attributes of a sample are extended to blocks to be estimated; and the concept of error estimation, where the validity of an estimation method is assessed based on the error involved [
11]. The methods utilized to perform the estimate are important, as they can influence the reliability and accuracy of the estimate. Several estimation methods have been proposed and implemented in the literature. These methods are largely categorized into geometric and geostatistical estimation methods [
11,
12], and they are termed conventional techniques in this paper. The geometric techniques (e.g., polygonal, triangular, random stratified grids, and cross-sectional methods) are simple and require few input parameters and are often applied at the early stages of a mineral project or to verify the results of the sophisticated estimation methods [
11]. However, the geostatistical (e.g., kriging, inverse distance weighting, and conditional simulation) techniques are more sophisticated. The conventional techniques have some inherent limitations. Most notably, they tend to perform poorly with highly heterogeneous datasets, overestimate or underestimate resources, and require significant manual processing [
18,
19]. Because of these shortfalls, currently, machine learning (ML) methods are being implemented in mineral resource estimation [
20,
21,
22,
23,
24].
ML techniques are algorithms capable of learning and modeling complex nonlinear patterns in a large dataset [
25,
26]. Since 1993, many authors have been exploring the potential of ML in resource estimation, resulting in several research publications. Even though the implementation of artificial intelligence and autonomous technologies in the mining industry began decades ago [
26,
27], it was not until 1993 that ML applications in mineral resource estimation gained enormous research interest. Zhang et al. [
28] noted that ML improves resource estimation in the following ways: (i) samples that are rejected in conventional resource estimates because they do not satisfy all quality control requirements can be used provided that the geological descriptions and measurements are reliable; and (ii) resource estimation block models can be constructed using fewer assays and more geology, leading to a reduction in operational costs. Additionally, the ML-based resource estimation approach is significantly cheaper and faster than conventional resource estimation [
28]. In addition, ML can modernize hypothesis-testing and geological modeling, contributing to the understanding of various deposit types estimation [
28]. Moreover, ML techniques can be employed to address operational challenges and improve safety in different sectors of the mining industry, including mineral prospecting and exploration, mineral evaluation, mine planning, mine scheduling, equipment selection, underground and surface equipment operation, drilling and blasting, mineral processing, and mine reclamation [
26,
29,
30,
31].
In addition to assessing the accuracy of different ML and conventional estimation models, these papers examine how various model variables, lithology, and data partition affect model performance. Many of these papers in the literature are spread across various journals and research databases. Therefore, it is imperative these works are consolidated, examined, and compared to form a coherent piece that would inform future research and serve as a reference for interested resource specialists. Our objective is that this paper would provide industry practitioners with an up-to-date knowledge about these emerging techniques and guide them on the choice of the technique applied for different estimation tasks. Additionally, it would also guide researchers to identify new ideas and areas requiring further scientific examination, as the review highlights some limitations and future trends of ML applications in mineral resource estimation.
The remaining part of this paper is divided into five sections.
Section 2 outlines the review methodology.
Section 3 discusses conventional resource estimation approaches, including geometric and geostatistical methods.
Section 4 reviews relevant machine learning techniques that are used for resource estimation.
Section 5 presents discussions and highlights key emerging issues that could be the focus of future studies.
Section 6 covers concluding remarks.
5. Discussion and Future Directions
The continued advancement in computer power has unraveled novel and sophisticated soft computing techniques capable of handling large and complex dimensional data structures that were not possible in the past. These soft computing techniques include machine learning coupled with big data technologies to create a new paradigm of powerful resource estimation models. Machine learning applications have infiltrated all fields and industries, from engineering to sociology, agriculture to astronomy. Some scholars even refer to it as the new electricity driven by data. Geoscience has also witnessed tremendous advances with the onset of ML. There is a consensus among geoscientists that ML algorithms are suitable for geospatial data that often exhibit complex and high spatial variations [
129,
130,
131], and the algorithms can produce superior results for classification and regression problems than geostatistical techniques, especially when the relationship is non-linear [
132,
133,
134,
135,
136,
137]. Some of the ML algorithms are universal, adaptive, non-linear, robust, and efficient [
129]. Another favorable characteristic of ML is that it allows the discovery of new information or a deeper understanding of existing geospatial data. This is evident in the assertion by Nwaila et al. [
138] that the advent of ML made sedimentological data, which were initially collected for qualitative assessment of gold mineralization, more meaningful and contextually relevant. ML-based estimation algorithms can accommodate a combination of several geospatial parameters for grade prediction and ore classification. In addition to the spatial location of drill holes and composite grade, lithological, geochemical, and rock imagery features, which are often overlooked or not accommodated in geostatistical techniques, can be included in the dependent variables of the model [
139], resulting in a more improved estimation model. The main appealing features of ML for resource estimation, in contrast to conventional estimation techniques, are that it (i) requires fewer data pre-processing, (ii) handles complex non-linear relationships, (iii) relatively cheaper and faster, and (iv) does not assume underlying spatial distribution and handles incomplete data.
Despite the strong potential of ML and its recent increasing application in the evaluation of mineral resources, geostatistical methods remain the benchmark estimation technique in the mining industry. ML techniques are considered more of a complementary tool for validating results obtained from geostatistical methods than sole estimation techniques for mineral projects. Glacken & Snowden [
36] noted that many estimation tasks are still conducted using polygonal techniques regardless of the availability of sophisticated estimation methods. The high preference for geostatistical techniques may be because they have been applied in the industry for a long period, forming the basis for many mine designs in operation today. In addition, their underlying framework of linear correlation of samples and stationarity [
97,
133,
140] makes them less computationally expensive compared to ML techniques. Furthermore, they can be integrated with other statistical concepts, for example, conditional simulation, which is a combination of kriging and the Monte Carlo sampling method [
36]. Moreover, many resource engineers have gained comprehensive experience over the years in applications of these methods with reliable results. Thus, they would be more comfortable working with techniques that are widely accepted in the industry. Stakeholders also have confidence in these techniques.
Additionally, most geostatistical algorithms have been automated in modern mineral resource software packages (e.g., Datamine, GEMS, Leapfrog, Micromine, Surpac, Vulcan, etc.) for easy data manipulation and estimation. This allows the estimators to focus more on the data preparation and result interpretation. Another reason could be that geostatistical techniques are widely recognized by the Committee for Mineral Reserves International Reporting Standards (CRIRSCO) members and their respective mineral resources and ore reserve estimation reporting schemes, such as National Instrument 43–101 in Canada, Joint Ore Reserves Committee Code (JORC Code) in Australia, South African Code for the Reporting of Mineral Resources and Mineral Reserves (SAMREC) in South Africa, SME Guide in the USA, NAEN Code in Russia, and Pan-European Reserves and Resources Reporting Committee (PERC) in Europe and the United Kingdom.
Although geostatistical techniques are the industry’s estimation standard, they possess certain inherent limitations. As mentioned earlier, these methods are generally linear estimators; thus, grades of unknown samples are determined based on second-order statistics and the use of linear correlation or variogram, which describes mineralization of a deposit [
97]. According to Das Goswami et al. [
134], second-order statistics work reasonably well with statistical processes following the Gaussian process. Nonetheless, it is difficult to model complex geological structures with high variability that do not follow the Gaussian process. Tutmez [
76] also pointed out that geostatistical methods perform poorly on small data sets (i.e., not enough data to achieve acceptable variogram calculation). Therefore, they are not suitable for small deposits or during the initial exploration stage, where data is limited. They also require significant data-processing [
137]. In such situations, ML techniques can be adopted.
ML techniques are known to sufficiently handle spatial uncertainty associated with geological data [
89,
97,
133,
141] sbecause of their ability to determine the relationship between complex and non-linear input and output variables [
97]. ML techniques can learn and map inherent relationships among the data variables. Given enough geological data (such as drill hole coordinates and assay results), ML techniques can learn the relationship between input parameters (coordinates of drilled holes) and output parameters (ore grades). The trained model can then be used to estimate grades for unknown points within the same geological area. In effect, no assumption is made about factors or relationships of ore grade spatial variations, such as linearity between drilled holes [
133] as in the case of geostatistical techniques. Several case studies (see
Table 3,
Table 4,
Table 5 and
Table 6) have illustrated the potential of ML techniques as a good estimator of ore grade for different mineral deposits; some studies have done so by comparing its performance with geostatistical techniques.
Results from studies comparing the performance of both models are mixed. In some cases, geostatistical techniques showed better output and vice versa. In other words, there is no outright best method between the two, and they seem to complement each other. Generally, the ML techniques seem to have higher accuracy than the geostatistical methods. When Das Goswami et al. [
97] compared two ML techniques (general regression neural network (GRNN) and multilayer perceptron neural network (MLPNN)) and one geostatistical technique (ordinary kriging (OK)) in an iron deposit, they found that GRNN exhibits better generalization potentials and also provides higher accurate prediction than MLPNN or OK models. MLPNN and OK were also observed to overestimate lower grades and underestimate the higher grades, while GRNN showed minor variation between the predicted and actual iron grade. In the Nome gold ore reserve estimation, Dutta et al. [
102] demonstrated that SVM produced better estimates compared to ANN and OK. Afeni et al. [
142] re-examined grade estimates for an iron ore deposit using MLPNN and OK and observed that the OK model showed superior performance than the MLPNN model. The total resource definition by OK was about 12% lower than that of the conventional method currently used in the mine. Karami and Afzal [
101] also evaluated the performance of ANN and IDW on a copper deposit and reported that the IDW method showed less variance, while ANN demonstrated high overestimation and underestimation. Thus, in this instance, IDW is a better estimator than ANN. In Kaplan and Topal’s [
73] study, the NN and kNN model underestimated any grades between the 15 ppm and 20 ppm range. Upon close examination of sample points, they observed that the network could not ignore the effect of discontinuity of lithology in areas where mineralization is structurally controlled and a test point located near a fault; thus, the model is more suitable for mineralization controlled by lithology than structure.
It is worth mentioning that ML techniques are not a panacea to all resource estimation problems, as they also have limitations. Jafrasteh et al. [
115] indicated that ML techniques are effective only if the training and test datasets have similar distributions. The models tend to perform poorly when the samples are far apart with increasing spatial variation. Thus, the training data and test dataset must be carefully partitioned to ensure that they have the same or similar geological characteristics. This problem can be resolved by assigning a higher weight to training samples closer to the test samples [
141]. Kapageridis [
143] observed that the dimension of input data influenced the performance of ML techniques, particularly ANN applied to ore grade estimation. After examining 2D and 3D spaces with varying input configurations, ranging from two to sixteen input dimensions on different deposit types (potash, marl, phosphate, and copper), the author concluded that for ANN, “there is no globally applicable configuration for all deposit and sampling scheme types, and each deposit and sampling scheme must be considered separately to find the best configuration applicable.” The adopted input dimension can either hurt performance when available data is small or improve when available data is large enough. Thus, it can be inferred that there is no one-fit-all ML technique for all resource estimation problems, but the choice of technique is contingent on available data and deposit characteristics. Further, we observed that many ML-based grade estimation models are limited only to geological parameters (e.g., using only spatial coordinates). Erdogan Erten et al. [
144] corroborate this observation, stating that “ML models provide useful tools to generate spatial estimations of geological features, but they do not consider the spatial dependence among the observations and they primarily use coordinates as predictors. Thus, many ML models produce visible artifacts in the resulting estimates along the coordinate directions.” The authors proposed using ensemble super learner (ESL) to address this weakness.
Again, ML algorithms that focus on finding the best solution for a model tend to overfit or underfit because the optimum model for training and testing the dataset may not necessarily have the best generalization ability [
145]. This problem can be addressed using ensemble methods, where multiple learning algorithms are applied with different parameters, and the resulting solutions are averaged to obtain better performance than any of the constituent learning algorithms [
146,
147,
148]. Examples of this learning method are Bayes optimal classifier, bootstrap aggregating (bagging), boosting, Bayesian model averaging, bucket of models, and stacking. Chatterjee et al. [
148] investigated the performance of genetic algorithms and k-means clustering based on an ensemble neural network for a lead-zinc deposit. The result showed no significant difference in the model performance. They attributed the marginal performance to the high coefficient of variations and skewness of the data sample. But Tahmasebi et al. [
78] pointed out that, perhaps, optimizing ANN’s parameters and topology could have improved the result. An earlier study conducted by Tahmasebi and Hezarkhani [
149] showed better results when the ANN’s parameters and topology were optimized with genetic algorithms.
Another critical factor observed to influence the accuracy of ML-based estimation models is data partitioning [
110,
150]. Most studies assume random partitioning to divide the sample data into training (e.g., 60%), validation (e.g., 20%), and testing (e.g., 20%). However, due to variation in drillhole samples (in both 2D and 3D) and erratic distribution of geochemical anomalies, such an approach may bias model performance as closer holes tend to exhibit similar features than farther holes. Thus, it is important to consider lithological features and other inherent characteristics particular to the formation when deciding which data partition regime to adopt. Additionally, instead of using composite values, individual drillholes can be modeled along the z-axis based on core sample intervals. The result obtained from each local drillhole model can be synthesized as input-output parameters for the global model. This would mimic spatial distribution along the drillholes, allow the inclusion of more features, produce more realistic models, and improve performance and accuracy.
ML is data-driven, requiring a large dataset for training, testing, and validation. The enormous volume of data needed to implement and derive meaningful results from ML successfully may not be readily available at the beginning of exploration programs or in greenfield projects. Acquisition of extra data is time-consuming and expensive. Therefore, ML may be more appropriate for brownfields projects, operating mines, or reassessing decommissioned projects. However, this is not to say that ML cannot be applied to small mineral estimation problems, as certain soft computing techniques such as fuzzy set theory [
74] are well suited for such problems. Other notable shortcomings of ML models include overfitting [
97,
132,
143] higher computational running time [
134], and smoothing effect [
132].
Considering the long usage of conventional estimation techniques, knowing that they have been tried, tested, and modified over the years to achieve the reliable estimates, it would be challenging for resource estimators to abandon these techniques for ML methods. Such a switch may result in chaos in the industry; stakeholders would doubt reserve estimates, and financial institutions would be reluctant to fund mineral projects because they may not trust the project value. In terms of resource estimation, ML techniques are not matured yet, and most of the applications are in academic settings. It would take time for it to get industry-wide recognition and acceptance. Studies should focus on integrating traditional/conventional and ML techniques and form hybrid algorithms to fasten the adoption process. It is worth noting that most ANN-based models in practice assume a deterministic approach to model non-linear systems [
151,
152]. However, natural phenomena such as the occurrence of mineral resources do not follow deterministic processes, as they are characterized by stochastic properties with high uncertainty. Consequently, Kaplan and Topal [
73] asserted that no matter how much network is trained to improve the prediction, independent stochastic events cannot be predicted by any neural network models. However, we believe that a more advanced ML algorithm such as stochastic neural networks [
152], which account for random processes in a system, and deep learning could address the heterogeneity of geological formations in resource estimation to some extent. These novel algorithms should utilize the merits of both conventional grade estimation methods and ML techniques and should also be able to estimate multiple ores since there are deposits with multiple mineralization, as many of the current ML-based mineral estimation models focus on only one commodity. Further, the ML and hybrid algorithms should be incorporated in mining software to make ML more accessible to resource engineers. Some scholars [
153,
154] are already building ML-based estimation software where the user would only provide sample assay data (e.g., borehole coordinates and ore grade), and the software will analyze and predict grade for unsampled areas.
The inherent heterogeneous nature of mineral deposits, which is evident in exploration datasets with varying mineralogy, mineral content, natural fracture, lithology, and other properties, can be likened to concept drift in machine learning. Concept drift refers to the unexpected changes in underlying data distribution over time [
155,
156,
157]. The concept suggests that as data evolve over time, the distribution underlying the data is likely to change, resulting in the poor and degrading performance of predictive models. Thus, the model must be updated regularly in real-time, as new data is obtained. The erratic distribution of ore grades means that resource estimates are subject to change as more data is gathered during the exploration or production phase. Therefore, ML-based resource estimation models should be able to analyze emerging data in real-time to reflect current grade distribution of a deposit. The idea is that the ML model should train, test, and validate each incoming dataset and use the result to update previous estimates. Studies have examined the problem of concept drift and proposed different approaches for detecting and handling it in several fields, including weather forecast, smart grid analysis, spam filtering, and predictive maintenance [
157,
158,
159,
160]. Žliobaitė et al. [
155] emphasized the application of Big Data Management and automation tools and the need to account for the evolving nature of data collected over time. Further, they recommended moving from the adaptive algorithms towards adaptive systems that would automate the full knowledge discovery process and scaling solutions to meet the computational challenges of Big Data applications. With Big Data applications, mineral exploration data can be organized in the form of data streams rather than static databases to enable online and real-time prediction. Zhukov et al. [
157] also proposed an approach using decision tree ensemble classification method based on the random forest (RF) algorithm for addressing concept drift in smart grid analysis. The proposed model compared well with concept drift approaches like Online Random Forest and Accuracy Weighted Ensemble (AWE). ML techniques such as lazy learning algorithm, incremental learning, support vector machines, classification trees C4.5, genetic algorithms, and neural networks have also been employed to address concept drift.
Such concepts can be extended to cover other downstream activities of mineral exploration and exploitation. For instance, after ore reserve estimation, ML techniques should be able to: develop block models based on the grade estimate, determine the cut-off grade, propose pit shells (geotechnical features of the formation included), perform project valuation [
161,
162], design mining sequence, and schedules for optimal extraction. All these processes can be automated as an integrated computer program, like automation of truck haulage, requiring less human interference and error. Full implementation of ML in resource estimation coupled with deep learning and other AI technologies like internet-of-things, drones, robotics, and blockchain would transform mining into the mine of the future being proposed by major stakeholders in the industry.
With the deployment of Big Data Management and Automation systems in the mining industry, we foresee the successful implementation of advanced ML techniques such as deep learning that requires enormous data to produce reasonable results. The availability of such datasets would help researchers apply ML techniques efficiently. ML applications in mineral resource estimation are currently focused on evaluating resources during the exploration stage, with little application post-exploration. We envisage that future studies would examine its applications during the operational stage of a mine, for example, grade control evaluation and reconciliation in surface and underground mining. In addition, given the recent proliferation of ML techniques, future research needs to determine which algorithms are most robust and appropriate for resource estimation. An industry-accepted ML application standard (ML Rubric) can be developed whereby all algorithms are subjected to a set of selection criteria, and every algorithm must obtain a pre-defined passing score before implementation. Some factors that can be considered in the ML Rubric are project goal, nature of data, ground condition, alteration levels, geological settings, model interpretability, minimum sample size, minimum features, acceptable error margin, performance metric, and computing resources.
Figure 12 illustrates a schematic of the ML Rubric factors, consisting of three main selection categories (algorithm attribute, project characteristics, and implementation process). Ultimately, such a standard would help harmonize ML-based models, eliminate discrepancies, and promote acceptance of ML applications not only in mineral resource estimation but other sectors of the mining industry as well.
It is important to indicate that a major drawback of ML applications is data limitation. The literature shows that most ML-based mineral resource estimation models were developed using composite values obtained from a few boreholes. Though these models produced satisfactory results, their performance can be improved with access to more data. The performance and accuracy of ML techniques, including deep learning, are highly dependent on a large and quality dataset that is partitioned appropriately into training, validation, and testing sets to ensure each set is a representative of the population [
31]. It is well known that the functional capabilities of artificial intelligence and ML rest upon a sufficient dataset. However, what constitutes an adequate dataset size is not well defined, as the amount of data required can vary from one project to another. In this regard, studies such as Ganguli et al. [
163] have provided recommendations to address ML application challenges peculiar to the mining industry. They recommended a comprehensive understanding of the modeling process before implementation and advised caution when using soft computing tools and software products [
31]. Their recommendation also included the random data partition in three subsets (i.e., training, testing, and validation). Further, they suggested that the training subset should contain the highest and lowest values, and samples should be assigned to the training subset first, followed by validation and testing, during data grouping/segmentation. Additionally, the best data collection and processing protocols should be observed during model development to minimize error and ensure the dataset is good enough for its intended use.