On the Challenges of Applying Machine Learning in Mineral Processing and Extractive Metallurgy

Abstract

The application of Machine Learning in Mineral Processing and Extractive Metallurgy has important benefits in terms of increasing the predictability and controllability of the processes, optimizing their performance, and improving maintenance. However, this application has significant implementation challenges. This paper analyzes these challenges and proposes ways of addressing them. Among the main identified challenges are data scarcity and the difficulty in characterizing abnormal events/conditions as well as modeling processes, which require the creative use of different learning paradigms as well as incorporating phenomenological models in the data analysis process, which can make the learning process more efficient. Other challenges are related to the need of developing reliable in-line sensors, adopting interoperability data models and tools, and implementing the continuous measurement of critical variables. Finally, the paper stresses the need for training of advanced human capital resources with the required skills to address these challenges.

1. Introduction

Machine learning (ML) is one of the most exciting areas in artificial intelligence, and “it is focused on teaching computers to learn from data and to improve with experience—instead of being explicitly programmed to do so.” [1]. After several decades of developing different learning techniques, ML is applied practically in all industries. In mineral processing and extractive metallurgy, the application of machine learning, particularly deep learning, has significant benefits. Among these, the following stand out [2,3]:

• the predictive modeling of processes, which allows for predicting their behavior, the variability of their outputs, and better control of these processes;
• the real-time analysis of material flow to optimize process performance through real-time analysis of operational variables;
• the predictive maintenance of equipment and components;
• the optimization of the use of energy and water.

In general terms, ML allows for optimizing process operation. For example, in the copper industry, “By ensuring that processing plants are consistently working in the upper range of their capabilities, machine learning can add 2 to 4 percent to met-al recoveries and 5 to 15 percent to throughput. Such improvements offer increased global production from existing and planned mines of half a million to one million metric tons of refined copper by 2032, creating $9 billion to $18 billion in value per annum across all sulfide concentrators” [4].

Nowadays, almost all mining companies are digitalizing their processes and applying different tools for analyzing the generated data [2]. These tools have received different names, such as big data analytics [3], machine learning [5], data science [6], or artificial intelligence [7], depending on the context or background of the researchers/engineers involved in the data analysis. However, in all cases, the main idea is to apply some learning algorithm to analyze the generated data to make decisions or assist the human decision-making process. For this reason, in this paper, we choose the use of the term machine learning.

The main challenge when applying machine learning in any application/problem, particularly in the case of mineral processing and extractive metallurgy, is the availability of useful data. This challenge is related to problems such as data acquisition and transmission in mining environments, the compatibility of the different computer systems used to store and analyze the data, the methods/paradigms employed to analyze the data, and the joint use of phenomenological and pure machine learning models.

By analyzing these different problems and how to tackle them, this paper expects to guide the practitioners and developers of machine learning solutions to mineral processing and extractive metallurgy. Section 2 addresses the general problem of applying ML in mining and metallurgical processes. In Section 3, the analysis is focused on mineral processing and extractive metallurgy. Finally, in Section 4, some conclusions are presented.

2. Challenges for the Application of ML in Mining and Metallurgical Processes

In general terms, the application of machine learning in industrial environments carries significant implementation challenges, such as the need for appropriate modeling and design of the projects, which must clearly identify the benefit to be obtained and the methodologies to be used; the scalability of the solutions to be implemented; the availability of the required human capital; and data privacy [3]. These challenges are encountered across all industries.

In the case of the mining industry, specific challenges are added, which are related to the characteristics of the mining business: operations located in geographically isolated places (desert, high mountains, etc.), difficulty of accessing sites (deep mines, etc.) and adverse environments for digital technology (presence of dust and humidity, abrasive materials, among others). In this context, the capture, storage, and transmission of the data to be analyzed take particular relevance [3]. Thus, the main challenge in mining is the availability of useful data. Useful data means not only having data for characterizing the different phenomena to be modeled but also human annotations, enabling the application of learning methods to these data (e.g., a human being is required to determine the condition, normal or abnormal, associated with the data obtained from a process or machine in a given situation).

The following subsections describe some of the main challenges of applying machine learning in mining. The emphasis is on the aspects related to data acquisition, transmission and usage, information and knowledge generation, and the methods used to analyze the data.

2.1. Data Acquisition, Transmission, and Usage

As mentioned above, the mining environment is generally hostile to electronic devices, and obtaining data through sensing is often complex. For example, in underground mining, the use of sensors inside the mine to observe environmental conditions such as humidity or geotechnical stability faces the problem that the sensors to be installed must be not only resistant to dust and humidity, but also to the conditions in the different unit processes such as blasting, which can damage these devices. In the case of metallurgical processes, many variables cannot be directly sensed, as the chemical processes involved can also damage the sensors. Therefore, an important area of current research and development is the development of sensors for mining and metallurgical applications that are robust enough to stand the operating conditions in which they will be used [3].

On the other hand, if the data captured are to be used for real-time decision-making, it is necessary to be able to transmit them effectively. One challenge to achieving this is the limited bandwidth of the communication networks used in underground and open-pit mining. These networks often operate at full capacity due to the requirements of transmitting video signals and control commands required by teleoperated or autonomous equipment operating under supervision. These challenges can be addressed by using higher capacity data networks (e.g., 5G) and communication protocols between sensors and data networks (industrial IoT), but also by using the edge computing paradigm, which seeks to perform most of the processing on the sensor itself, to minimize the amount of data to be transmitted, by extracting the relevant information in situ [8].

A third challenge relates to data ownership [9]. This situation occurs in the case of vehicles and large mining equipment, which have many internal sensors and therefore generate a significant amount of data (e.g., vital signs). However, many purchase contracts for the equipment establish that the data generated by the sensors (when the plant enters into operation) cannot be used by the company that owns the equipment for analysis and decision-making (for example, in predictive maintenance applications) as it belongs to the manufacturer. Naturally, this challenge is not of a technical nature but contractual, but nonetheless, it must be addressed.

A final challenge concerns data format and compatibility. It is necessary to have interoperable systems based on standard data formats. There are initiatives such as the “Technology program for the creation and adoption of international standards for mining interoperability” of the Chilean Alta Ley Corporation [10] and the Global Mining Guidelines Group (GMG) [11] that are addressing this issue.

2.2. Data, Information, and Knowledge

Several studies indicate that a significant part of the data generated in the different mining and metallurgical processes is not used in decision-making and that it is not even stored. For example, in Durrant-Whyte et al. [12], it is pointed out that less than 1% of the data generated in mining and metallurgical processes would be used in decision-making processes. Naturally, considering the great variety of mining and metallurgical operations that exist in the world and their different unit processes, it is very difficult to estimate the percentage of data effectively used. We can say with certainty that the fraction of data generated in mining and metallurgical operations and used in decision-making is low. This situation must be reversed if the benefits of digitization are to be fully exploited.

On the other hand, the concept of data is usually confused with that of information. It is forgotten that data can be redundant, correlated, or provide no information. The information corresponds to “data that have been processed in such a way as to be useful” [13]. Information Theory provides formal methodologies that allow for quantifying the amount of information contained in the data. In fact, using these methodologies, data can be compressed without losing information. A significant challenge is using algorithms that allow for extracting relevant information from the data available at different stages of the processing pipeline, to store, transmit, and analyze only the data containing relevant information.

Knowledge is generated from data and information, allowing the understanding of data patterns and generating models. In many applications, the generation of predictive models enables automated decision-making or assists human decision-making processes. However, developing useful models requires a phenomenological understanding of the modeled processes. Hence, the challenge for generating these models is the effective collaboration between specialists in machine learning and professionals who understand the different processes to be modeled (metallurgical engineers, mining engineers, hydraulic engineers, geologists, etc.).

2.3. ML Methods and Paradigms

The machine learning methodology to be applied depends on the problem being addressed (classification, regression, estimation, clustering, fault detection, modeling, control) and on the quality and quantity of the available data. This is a relevant issue in mining and metallurgical applications where data scarcity is frequent.

The supervised learning paradigm is applied when a statistical classifier is trained using available data in the form of feature variables (the input of the classifier) and their associated labels (the output of the classifier) [14]. Then, having a normally large number of examples, i.e., a set of examples of the input-output variables, the statistical classifier approximates the function that relates the input with the output variables. This paradigm is typically used to address classification, regression, estimation, and control problems. However, the paradigm cannot be used when many examples are difficult to obtain. For instance, powerful algorithms, such as Deep Learning algorithms, usually need many examples [15]. Reasons for not having a large number of examples are data scarcity, for instance, due to the complexity of obtaining sensor data from a mining process, or the high cost of human-based labeling of the examples, or because a condition that needs to be predicted occurs very infrequently (e.g., a SAG mill failure).

The application of the unsupervised learning paradigm [14] does not require labeled data because it is based on finding relationships between the examples using a self-supervised or self-organized process. Clustering algorithms and dimension reduction algorithms, such as Principal Component Analysis [14], are examples of unsupervised learning.

Many efforts are being devoted to using unsupervised learning to train statistical classifiers. For instance, contrastive learning [16] looks to learn low-dimensional representations of data examples by contrasting/comparing similar and dissimilar examples. The use of auto-encoders for training the convolutional layers of a Convolutional Neural Network [15] is an example of the application of this paradigm. Thus, the application of contrastive learning allows for using unsupervised learning for training the parts of the classifiers that extract the features and supervised learning for training the algorithm in charge of the classification task. This reduces the required number of labeled data. The semi-supervised learning paradigm [17] combines supervised and unsupervised learning to train a statistical classifier. Under this paradigm, the training is carried out using a small amount of labeled data and a large amount of unlabeled data. Both contrastive learning approaches and semi-supervised learning approaches allow for addressing the data scarcity problem produced in cases where the labeling process is challenging to implement.

The reinforcement learning paradigm [18] is based on using rewards to train an agent/controller that interacts with its environment and executes actions/behaviors. This allows learning through a trial-and-error process. This paradigm can address different problems, including learning controllers for metallurgical processes that are difficult to model (e.g., heap leach piles).

For predicting abnormal events or conditions (e.g., equipment failure), the typical ML approach is to characterize the normal condition using a statistical distribution model and then to detect deviations from this model. A deviation from the learned model will correspond to a detected abnormal condition/event.

Moreover, the use of phenomenological models, which from the ML point of view correspond to a priori knowledge, can speed up the learning process or require a smaller data set. For example, using a model-based reinforcement-learning approach can allow for learning to control a process using a smaller number of training steps (e.g., interactions with the environment) than when using a model-free approach.

In summary, the complexity of the mining and metallurgical environments, where frequent problems are data scarcity and the difficulty in characterizing abnormal events/conditions as well as modeling processes, requires the creative use of different learning paradigms (supervised-, unsupervised-, contrastive-, semi-supervised-, reinforcement-learning, among others). In addition, using ML methods that use phenomenological models can make the learning process more efficient. The next section addresses these challenges by analyzing relevant issues such as: (i) the data collection and sampling process, and the need for developing reliable in-line sensors and continuous measurements of critical variables; (ii) the need for appropriate data analysis tools which can interoperate with the current software tools used in the control systems of mineral processing and metallurgical plants, (iii) the predictive modeling of processes, and (iv) future challenges and trends.

3. ML in Mineral Processing and Extractive Metallurgy

This section describes the current status of the use and future challenges of ML application in mineral processing and extractive metallurgy, particularly in comminution, flotation, and hydrometallurgy. In this last case, the analysis focuses on heap leaching operations. Even though smelting and refining operations have been excluded from this study, several aspects such as lack of sensors, a difficult sampling strategy, or data interpretation are similar to comminution, flotation, and hydrometallurgy. However, the batch configuration of smelting and refining processes defines additional challenges.

Thus, the use of ML in mineral processing and extractive metallurgy can consider steps of data collection, particularly associated with sampling and measurement, data analysis, and finally, prediction [19] (see Figure 1).

In this regard, the discussion of this section is based on each step defined in Figure 1, considering a review of the literature and industrial experience [3].

3.1. Data Collection: Sampling and Measurement

The control and maintenance of mineral processing and extractive metallurgy plants (i.e., comminution, flotation, and hydrometallurgy operations) are among the most critical in the mining industry as these stages transform the ore rock into a valuable product. However, these operational units are also the most intensive regarding resources such as water and energy. For example, the comminution stage alone can represent up to 50% of the energy consumption of a mining operation and, therefore, a large part of the operational cost [20].

Accordingly, the main objective of process control is to maximize the plant operation efficiency at a minimum cost and resource expenditure [21,22]. The operational performance depends on relevant variables for each process, such as the objective grinding size (P80), which would liberate the mineral of interest; the acid consumption in the copper leaching stage; the flotation reagents consumption; and also, the tonnage required to be processed. Therefore, monitoring a plant and the feed is relevant to maintaining and improving operational performance. For instance, parameters such as the particle size distribution, the mass flow, and the grade of the element of interest or the density are necessarily to be quantified through instrumentation in the plant. Generally, these measurements are not obtained with great precision but rather from a large volume of data, with a recurrence that ensures stable and continuous operation (e.g., minute-by-minute or even second-by-second).

Currently, magnetic flow meters, nuclear densimeters, X-ray analyzers, adaptive controllers, tank level sensors and pipe pressure gauges, among other equipment, are used to monitor and control each flow, from comminution to the generation of tailings [21,23]. These sensors generate a large volume of data, up to several gigabytes of information per minute. Yet, due to the complexity of the mining-metallurgical processes, some critical measurements cannot be carried out with enough precision with the existing instruments, or it is impossible to measure some of the variables. For example, some variables that cannot be measured online are used to determine the efficiency and productivity of a process plant, such as flotation or leaching recovery, reagent consumption (acid in the case of leaching, foaming agents and collectors in the case of flotation) and the concentration of elements in leaching solutions. Therefore, these need to be measured by sampling the streams. Other important parameters to be considered when sampling the plant are the ore characteristics such as hardness, degree of liberation, ore grade, and mineralogy. These characteristics are determined by a laboratory after several hours and even days from the sampling process, which is performed daily or in each shift. The use of the information generated by continuous (instruments) or discontinuous (sampling) measurements has different objectives. The first is generally used for process control to ensure operational continuity and not necessarily to achieve optimum performance conditions.

Meanwhile, the discontinuous information is used to report productivity and take measures for process improvements, unfortunately in a reactive manner. Indeed, results from sampling are commonly used to quantify process efficiencies or to perform metallurgical balance for production. These limitations do not allow the development of prediction tools or control methodologies that optimize the process globally. For this reason, it is relevant to make progress towards developing measurement systems suitable for the different processes or requirements and more robust data-processing systems and prediction of results.

The reduction in head grades and more complex mineralogy of deposits, alongside the advancement of new technologies, have motivated the implementation of recent developments in terms of instrumentation and predictive models. An example in the field of ore characterization is the incorporation of high-resolution images to characterize the material that enters the plant. One of the problems in mineral processing is not knowing how the material will behave in the concentrator. Systems such as VisioRock [24] or WipFrag [25] allow for controlling cameras located on the conveyor belt, in addition to connecting to the control system to provide information about the size of the feed and report changes in the rock. Another recent development is commercial equipment that can be installed in the process (usually conveyor belts) for elemental analysis—X-ray diffraction spectroscopy (XRF) and laser-induced breakdown spectroscopy (LIBS) [26]. However, the presence of this equipment is not yet massive in mining operations, perhaps due to cost or simply because it is not considered in the design stages.

On the other hand, online mineralogy analysis is still in the research phases, but it has moved forward in the last few years. In this regard, interesting advances are using hyper-spectral analysis [27,28]. However, there is still a lack of development for its implementation in actual conditions, where the mineral is exposed to movement, dust, and humidity, among other factors. Understanding how geology controls rock behavior in processing is still a subject of much research, which requires detailed characterization.

Heap leaching processes are typically developed over months or years, depending on the ore mineralogy and metal of interest, determining large irrigation areas and heap heights ranging between 3 and 70 m [29]. In this context, the available information to control and operate the process is critical. These huge reactors require the measurement of several variables to ensure correct control of the process, such as moisture, metal concentration (e.g., copper, gold, uranium, among others), relevant cations (e.g., ferric ion, ferrous ion), relevant anions (e.g., chloride, sulfate, nitrate), temperature, dissolved oxygen, pH, oxide-reduction potential (ORP), and intrinsic permeability, among others [30]. However, there is a lack of instruments or sensors commercially available to measure these variables in real-time and in different places in the heap, particularly inside the ore bed. Unfortunately, this limitation of the information available restricts the possibility of improving the heap leaching operations, and the development of sophisticated models to predict metal production is typically limited by the information required [31], restricting the use of empirical models with low predictive capacity and, thereby, high uncertainty.

One of the most critical limitations in hydrometallurgical processes is the real-time and automatic measurement of the feed ore mineralogy, the metal grade of the feed ore and leached tails, and the metal concentration in solutions. The industrial state of the art considers sampling once or twice daily in relevant process stages, limiting the possibility of implementing short-term predictive or even reactive control. The measurement of ion content in heap leaching solutions is still carried out manually, procuring discrete data, as there are no reliable instruments commercially available to work with a multi-element solution. This fact limits the predictive control in hydrometallurgical plants. Even though the heap leaching process is typically slow, so the metal concentration in the pregnant leach solution (PLS) does not vary dramatically daily, the lack of instruments limits the measurement of the effluent quality in different places on the heap.

Particular to heap leaching, there are additional critical variables, in addition to other conventional leaching processes, such as intrinsic permeability and bed moisture [29], which influence the result of the leaching process. For example, the intrinsic permeability of the bed indirectly defines the irrigation flow rate in the process. On the other hand, the bed moisture establishes a pile’s geotechnical stability and the efficiency of the irrigation. Indeed, there are different developments for measuring the moisture in heap leaching processes. One of the techniques used is electrical resistivity, which is already a commercial technology for mapping 3D moisture in a heap pad [32,33]. Although these are still discrete data of a specific time in the process, which limit their applicability for operational control, it is quite helpful to be able to characterize zones with inadequate irrigation or under risk due to moisture values which are near saturation. Recently, thermal imaging analysis with unmanned aerial vehicles (UAVs) or drone-based remote sensing has been used to measure the moisture, saturation zones, and irrigation quality on a heap-pad surface [34,35,36].

Moreover, drone-based hyperspectral remote sensing was studied to determine mineralogical characteristics in an actual heap leach pad at the Safford copper mine [37]. These technologies could gather data in real time, so their implementation with a control system could be possible. Even though drone-based remote sensing cannot measure the internal characteristic of a heap pad, the data collected on the surface could be correlated to other variables of the heap leaching process.

Indeed, there are differences in the data captured within different mining operations. Each site’s standard of instrumentation fundamentally determines these differences. This means that the capacity to generate information and perform data analysis depends on the instruments installed in each operation. For example, in some operations, some instruments generate a large quantity of data—some relevant and others not so much—but are used for the registration or automatic control of minor equipment and accessories such as pumps. This information guarantees operational continuity without aiming at maximizing process efficiency. This operational standard is due, in part, to the lack of online instruments or continuous measurement of critical variables, such as mineralogy and, particularly, in the characterization of aqueous solutions.

Although the data generated from the online instruments and sampling can accumulate a large amount of information, data treatment in mining-metallurgical processes is far from other industries. Today, handling large amounts of data and defining how to use it are still some of the biggest challenges when using integrated platforms. The manipulation of a large volume of data, for example that carried out by digital technology companies such as Google, Facebook, or Amazon, is not a reality yet in mining. Indeed, the absence of phenomenological models that can handle these large datasets and their reliability to generate short-term responses limits the use of data, where ML methods to appear to be an appropriate solution.

In summary, the most relevant aspects of the current use of data for ML in mineral processing and extractive metallurgy are as follows:

• The current plants have different characteristics for measurement, recording, and control of variables, which depend fundamentally on the standard and size of each plant.
• There are critical variables in mineral processing, such as the mineralogical characterization, the particle size distribution, the metal grade in the ore, and the concentration of ions in solutions, which are not able to be measured in real-time due to the lack of accuracy in instruments that can operate under the real conditions of a plant.
• The generation of real-time data on some variables, such as flow rates, temperature, current, or pressure, must be related to discrete data in the same time period. The difference in the source and type of data restricts the use of predictive models.
• One of the most relevant challenges for applying ML in mineral processing and extractive metallurgy is the development of new sensors or instruments for acquiring the critical variables of the different processes.

3.2. Data Analysis

Most parts of mineral processing and metallurgical plants exhibit a certain level of control system (Process Intelligence—PI) that stores the operational information and assists in automating such operations. These systems aim to gather recurring information to take effective actions toward securing the operation’s stability [22]. The roles of PI Systems can be summarized as follows: to ensure system-operator interaction, acquire and process plant information to maintain its regular operation, save information from connected machines and make it available to metallurgists.

A PI system can also be associated with an expert control system, which can be integrated into the same platform, such as FLSmidth’s Outotec ACT or ECS [38], and uses algorithms to establish relationships between the variables coming from the different equipment in order to manage the plant. The OSIsoft PI System is widely used in the industry, allowing the automatic and real-time collection and analysis of data from the plant. Additionally, it allows interaction with the operator, reporting key performance Indicators (KPI values) and promoting plant automation [39]. Figure 2 shows an example of data stored by a PI System in a froth flotation process, showing a subset of the multiple variables that can be extracted and analyzed. That is the information metallurgists observe without any further detail. Indeed, for instance, in this plant, the molybdenum selective flotation separation is performed using nitrogen gas coming from a pumping system that filters the air. Increments in the nitrogen flowrate will inevitably lead to inefficiencies in air filtration incorporating more oxygen gas into the flotation cells. The latter will reduce the efficacy of NaHS towards decreasing the floatability of copper sulfides as oxidation mechanisms will be triggered. In the figure, after 11-03 in the time axis, the strong increase in nitrogen gas flowrate (N2 flowrate) and its subsequent fluctuation do not correlate as may be expected with NaHS dosing (NaHS rougher). Many arguments may explain this behavior, such as that changes in the gas flowrate do not immediately translate into the sudden oxidation of NaHS so increments in the reagent dosage are not required, or that changes in one variable will produce variations in other parameters that should be treated stochastically rather than deterministically only by defining a time delay. Such tough interpretations and decisions are those that professionals who deal with these PI Systems are required to undertake on a daily basis.

Regardless of the chosen method of collecting information (continuous or discrete data), finding operating conditions that maximize the efficiency of a process is tough to achieve as the previous data are biased only by the operators’ best judgment [41]. A typical data analysis pipeline in a mineral processing or extractive metallurgy plant is shown below:

• A first stage which involves implementing a set of simplified procedures that seek to clean up the data by removing impractical data (data conditioning/cleaning) is commonly implemented. This should not be confused with the simple removal of outliers. It is, in fact, more than that; it deals with real anomalies within a plant coming from unadjusted or poorly maintained sensors. Excessive data cleaning, though, can easily lead to studying biased scenarios based on erroneous judgments. Another challenge related to language is involved in the nomenclature used to indicate effects and causalities within the plant data which is highly variable, and in many cases is not rigorous.
• Spatiotemporal traceability and consistency between the results coming from different measurements. Attempting to incorporate the temporary differences between the various measurements into the analysis would provide a better understanding of the impact of the variations of the parameters and the causality that these generate. One of the most relevant problems behind this issue is that the exact composition mineral-wise is unknown in a concentrator plant. Geometallurgically-based strategies (assumed as proxies of operations) have proved to be of great value in this area, significantly reducing the risks behind plant operation [42]. However, they are not the unique solution to all problems. An excellent correlation to all events—among parameters derived from intrinsic measurements of the mineral when it is processed—has not allowed the production of generalized models that could accurately predict the efficiency of the process. Interestingly, precision is one aspect that has been achieved in mining operations, but in general terms, it is only a secondary derivative of the actual problem.
• Statistical analysis of the data. According to the methodologies used in ML, the definition of statistics is more general than simple data recording and analysis, looking for a rather holistic approach where the life cycle of the data should be considered. How the data will be used and where they will be delivered depends on where the data come from. For instance, in the case of froth flotation, it is easy to indicate where the data come from beyond the challenges involved in determining their representativeness and reproducibility. Due to the lack of predictive fundamental models, the data are currently only used to build multivariable models and simple relationships with common efficiency outcomes such as recovery and grade [43]. On the other hand, where these data go is a challenge that has not been addressed in detail. The metallurgical results, beyond generating indications of production, are not shared between the different departments taking part in the business sequence. Local decision-making based on marked information (equivalent to what can be a trending topic) is carried out periodically by operators and metallurgists. The challenge behind this approach is that it is unclear when these decisions should be translated into actions. At a mining site, data generated internally from a PI System are barely used for a decision-making process. To exemplify this, a process improvement of the plant usually involves a metallurgical sampling campaign or plant survey, a strategy that has been used since the 1960s [44]. This action requires a plant in a steady state condition, since the most widely used models only address mass balances in that condition. In this scenario, the use of ML could be observed as less attractive. When evaluating whether a plant is in a steady state or not, only some streams and equipment within the plant are verified, such as checking whether a flotation tank is not overflowing abnormally. However, in most cases, there is no standard indicating the stability of the plant. Furthermore, variables chosen to build models such as reagent consumption and liberation still represent challenges regarding real-time measuring techniques, and new sensors and analytical techniques are still to be implemented. For example, hydrophobicity, a key aspect of froth flotation, is yet somehow troublesome for many metallurgists. Enormous efforts are made by some mining operations when implementing information integration units or setting up training courses that allow the different areas of the business to speak a common language and understand the relevance of the information universally in mining. Such integration of information coming from different sources and their correct interpretation will be crucial for ML success and, therefore, to build models that could be more responsive to the dynamic conditions of the plant.

3.3. Data as Input for Prediction: Predictive Models

Concerning modeling using the characterized data, the tools are used to reconcile geological information and its response at the metallurgical level. The main objective of these methodologies is to be able to generate predictive responses in the plant based on characterization and, therefore, to be able to feed a mine-planning model with more accurate information. However, in order to generate predictive tools, a large amount of information from metallurgical tests that should be associated with geological characterization is required. For reasons of cost, this is impracticable, meaning that the metallurgical information is usually lacking. Therefore, operating plants have implemented ML methodologies based on historical responses of minerals in the plant. This could be improved by a better understanding of the effect of geology on the processes that explain some plant variations [45].

On the other hand, the efficiency of an operating plant also depends on the smooth and continued operation of the equipment, which needs to be close to the design capacity. An essential condition for the equipment to operate efficiently is the maintenance of the components. Preventive maintenance before sections wear out can ensure a stable throughput in the comminution plant. For example, the early replacement of liners and preventive maintenance of the pinion gear of the mill can avoid a diminution in the grinding capacity induced by the maximum height of the load and the rotation speed. The loss of equipment and a full production stop can economically hinder the mine operation more than short stops for keeping components healthy [46]. However, identifying when the piece should be replaced is difficult as it depends on factors such as previous operating conditions and the ore’s hardness and abrasive characteristics. The data collected from mills, such as the shell’s vibration or acoustic emissions from installed sensors [47], can indicate the operational condition. ML tools can be applied to this data to implement maintenance advice in comminution plants to keep their stability [46].

Regarding predictive models for heap leaching operations, the most used for design and operation are empirical-based models [31] that typically use metal extraction curves obtained from small-scale column leaching tests. These models have high predictive uncertainty. On the other hand, phenomenological-based models have been developed for the bioleaching of copper sulfide ores [48,49] with high predictive capacity but high requirements of available data from the operation or lab-scale tests to input in the models. Other models and approaches have been studied for different purposes in heap leaching processes, such as the use of artificial intelligence and machine learning tools, for example random forest [50], artificial neural networks, and support vector machines [51] for improving the prediction of copper production, decision support systems for bioleaching processes [52] or acid leaching for oxide copper ores and chloride leaching for copper sulfide ores [53], stochastic models based on Bayesian networks [54], analytical models [55] and optimization methods [56]. The main limitation for predictive purposes of these models is the requirement for real data from the plant. However, their implementation is associated with a PI system, where the generated data can continuously be used by each model to improve the operational results in the heap leaching process [4].

Regarding the general problems of controlling metallurgical processes of different types, the standard control system approach, which can be implemented using classical or data-based controllers, requires having a model of the dynamics of the process to be controlled. Thus, when the dynamics of the process are difficult to model, this methodology cannot be applied (e.g., the real-time dynamics of a heap leach pile are extremely difficult to model). As an alternative, new paradigms that are not based on the classical control system approach, such as reinforcement learning [17], can be used.

3.4. Future Challenges and Trends

Summarizing, the most relevant advances in the use of ML in mineral processing and extractive metallurgy are the following:

• There are useful developments of new technologies related to the measurement of online variables, particularly for mineral characterization, based on X-ray, imaging analysis and processing, and hyperspectral or laser image analysis. In addition, using drone-based remote sensing for mapping the moisture on the surface of the leaching pad could be an interesting application if these data can be correlated with variables inside the heap bed.

  ▪

  Implement elements from the Internet of Things (IoT) or Industry 4.0, such as wireless data transmission.

  ▪

  Standardize data and develop interoperability data models and tools. Then, these models could interact with the information generated by the instruments from the plant and respond in a limited time to be integrated into a plant’s control system.

  ▪

  Implement the continuous measurement of critical variables of processes in a way that is currently not used, such as ore hardness, particle size distribution, moisture, intrinsic permeability, acid consumption in heap leach plants, and flotation reagents, among others.

• Advances in empirical, phenomenological, and pure ML-based models for different processes or unit operations are allowing a better understanding and decision support to operations. However, several issues regarding data requirements and predictive accuracy still need to be improved. The combined use of phenomenological and ML-based models can take advantage of the benefits of each one to obtain better modeling, control, or decision-making. In this regard, it is necessary to improve the models of different unit operations involved in a plant (or develop new ones), such as crushing, grinding, flotation, or leaching, to enhance the reliability and predictability of results under different operational conditions or ore characteristics, including data with high variability. Hence the use of joint phenomenological and ML-based models, to take advantage of the expert knowledge of the processes and the capabilities of the ML to better model and control the different processes.
• Improvements in methodologies of geo-metallurgy to allow the fluid connection between the geological information and the processing plant and their impact on the mine planning models.
• In cases where the available data are not sufficient to apply the supervised learning paradigm, use alternatives such as unsupervised learning, contrastive, semi-supervised, or self-supervised learning, reinforcement learning, and prognosis tools can be useful.
• The use of ML algorithms that are not based on the supervised learning paradigm will allow for addressing applications with modeling problems, e.g., the process is challenging to model, data scarcity problems, sensor measurements are difficult to obtain, or where the labeling process is expensive.

4. Conclusions

The article has shown that using ML in mineral processing and extractive metallurgy can increase the predictability and controllability of the processes. Nevertheless, there are specific challenges that must be addressed before applying ML in mineral processing and extractive metallurgy plants aiming to enhance operational efficiency. The main challenges are related to reliable and accurate sensors for critical variables in the plant, the improvement and predictability of mathematical models, both phenomenological-based and empirical-based, and the implementations of geo-metallurgical methodologies.

Therefore, the digital revolution in mineral processing and extractive metallurgy plants, specifically the increasing use of the ML paradigm, will promote a change in the workforce, where the specialties associated with data analysis, software development, and digital design will be of great relevance to test the new developments, analyzing a huge amount of data, and working in systems that will be moving forward to the IoT. Therefore, training advanced human resources with skills in these areas will be relevant to overcome the challenges of mineral processing and extractive metallurgy, considering the metal grade decline, more complex ores, and the multi-metallic requirements of the technological industry.