Evolution of Research Stages for Artificial Intelligence and Machine Learning in Mineral Exploration

1. The Initial Stage: Laying the Foundation for Quantitative Prediction Paradigm (1970-1999)

The 1970s to 1990s were a crucial period for the construction of a quantitative methodology system for mineral exploration. Although a machine learning technology framework had not yet been established during this period, the innovative application of statistical models such as logistic regression and the weights of evidence method enabled the first mathematical integration and spatial correlation analysis of multi-source mineralization information, laying the scientific foundation for the standardization of exploration data and model-driven prediction paradigms. While neural network technology was still in the experimental exploration stage, its early attempts revealed the potential of nonlinear modeling, becoming an important precursor to subsequent breakthroughs in intelligent algorithms.

The systematic application of mathematical methods in geology during this period completely transformed the traditional experience-driven exploration model. Since the International Association for Mathematical Geosciences (IAMG) established six recommended methods for quantitative prediction of mineral resources in 1978 (Cargill and Clark, 1978), multivariate mathematical models have been deeply integrated into the quantitative prediction and evaluation of mineral resources by the USGS, the Canadian Geological Survey, and China. Among them, the evidence right method has become a core tool due to its advantage of integrating multi-source data. Agterberg's (1989) research published in Science improved the scientific validity of mineral exploration target delineation by integrating geological, geophysical and geochemical data through Bayesian probabilistic models (Agterberg, 1989). With the introduction of fuzzy logic (Cheng, 1996; Cheng and Agterberg, 1999), asymmetric metrics and uncertainty modeling techniques have driven the transformation of mineral exploration prediction from a deterministic to a probabilistic paradigm. This has spurred the development of quantitative resource prediction and evaluation methods, including the right of evidence method (Bonham-Carter et al., 1989), fuzzy right of evidence (Cheng and Agterberg, 1999), logistic regression (Agterberg and Bonham-Carter, 1999), fuzzy logic (An et al., 1991), and the evidence trust function (An et al., 1992), as well as methods initiated by Chinese scholars in 1990 such as "Geological Anomaly-Induced Mineralization and Metallogenic Prediction" (Zhao Pengda and Chi Shundu, 1991) and "Comprehensive Information Metallogenic Prediction" (Wang Shicheng et al., 2000), which Chinese scholars began developing in 1990 (Zhao Pengda, 2002).

During this period, a mathematical framework for expressing geochemical sampling and geological mapping data was simultaneously constructed, promoting the theoretical foundation for the standardization of exploration data. The systematic application of mathematical methods propelled the transformation of geochemical data from qualitative description to quantitative analysis, establishing standardized data processing workflows (Cargill and Clark, 1978). In the field of geological mapping, mathematical models enabled a paradigm shift from manual mapping to digital spatial representation, significantly improving the machine operability and analytical efficiency of data.

Meanwhile, neural network technology saw early exploration. Despite limitations imposed by computing power constraints in the 1990s, feedforward neural networks were pioneered in mineral prospect mapping (Singer and Kouda, 1996; Clare et al., 1997; Singer and Kouda, 1999). Although large-scale application was not achieved due to hardware limitations, these studies were the first to verify the feasibility of nonlinear algorithms for processing complex geological systems, laying the experimental foundation for methodological breakthroughs in the era of deep learning.

2 Preliminary Exploration of Machine Learning (2000-2015)

From the late 20th to the early 21st century, Artificial Neural Networks (ANNs) began to be applied in the field of mineral exploration (Brown et al., 2000; Harris et al., 2003; Rigol-Sanchez et al., 2003; Porwal et al., 2004; Behnia, 2007; Nykänen, 2008). The foundational research by Brown et al. (2000) first coupled artificial neural networks with Geographic Information Systems (GIS) to achieve integrated prediction of multi-source regional data. Based on a 1:100,000 scale rasterized database of the Tenterfield region in New South Wales, the study evaluated the performance differences between ANNs and the weighted evidence method in predicting gold mineralization favorability. Their core findings showed that the ANN model not only significantly outperformed the weighted evidence method model in prediction accuracy, but this superior performance was achieved with only about 63% of the training data required by the weighted evidence method. However, the practical application of ANNs at that time faced significant bottlenecks: on the one hand, limited by the computing power and available data scale, model building relied on manual feature engineering by researchers, with a single model training session often taking several weeks. On the other hand, the inherent "black box" nature of the model made it difficult to explain the geological mechanisms underlying the predictions. This lack of core interpretability made it difficult for the predictions to effectively guide exploration decisions, thus severely restricting the widespread adoption of ANN technology in practical exploration at that stage.

During this period, supervised learning algorithms achieved breakthroughs in mineral prediction (2000s). Support Vector Machines (SVM) and Random Forests (RF) gradually replaced traditional methods as mainstream algorithms, marking a paradigm shift in machine learning from theoretical exploration to engineering practice. In SVM applications, Zuo and Carranza (2011) constructed a gold mineralization probability spatial model by integrating multi-source geological variables from the Meguma gold deposit in Nova Scotia, Canada. Their prediction results reduced the total error of mineral prospect classification by 5%–9% compared to the weight of evidence method, demonstrating for the first time the significant advantage of SVM in multi-dimensional evidence layer fusion. In RF applications, RF, due to its unique anti-overfitting properties and noise tolerance, performs exceptionally well in small-sample data scenarios: Carranza and Laborte (2015) established a high-precision prediction model based on training samples from only 19 hydrothermal gold deposits in the Baguio region of the Philippines, providing a new approach for exploration in areas with scarce small-sample deposit data. A comparative study by Cracknell and Reading (2014) further showed that RF has the best overall performance in intelligent geological body identification tasks. However, this stage still faces dual constraints: firstly, model training heavily relies on high-performance computing clusters, making real-time deployment in field exploration difficult; secondly, the model's cross-regional generalization ability is significantly weakened due to the limitations of historical data scale and distribution, restricting the technology's universality. It is worth emphasizing that the feature engineering experience and data standardization paradigm formed during this stage laid an indispensable technical foundation for the subsequent development of deep learning technology.

3 Rapid Development Period – Deep Learning and Big Data Driven (2016-2025)

With the breakthrough in the heterogeneous computing power of GPUs (Graphics Processing Units) and the exponential growth in the volume of multi-source exploration data (hyperspectral remote sensing, 3D seismic data, UAV aeromagnetic data, etc.), deep learning-driven mineral exploration has entered the intelligent decision-making stage. The core feature of this paradigm shift is that "data-driven feature learning" replaces the "manual feature engineering" of traditional mathematical models, significantly improving the ability to represent complex mineralization patterns. However, this technical path faces a dual challenge: firstly, the lack of inherent interpretability of deep neural network models leads to ambiguity in the geological genesis mechanisms of prediction results, restricting the reliability of exploration risk assessment; secondly, the real-time data closed-loop enabled by edge computing requires algorithms to complete tasks such as real-time identification of exploration targets, dynamic interpretation of mineralization anomalies, and drilling path optimization in resource-constrained environments, placing higher demands on model lightweighting and embedded deployment. Current research is focusing on "deep learning + real-time data processing," using graph convolutional networks to process the topological relationships of 3D geological bodies, knowledge distillation to achieve lightweight model transfer, and attention mechanisms to improve the interpretability of anomaly detection, thus driving a leapfrog evolution in exploration decision-making from static prediction to real-time feedback.

Bergen et al. (2019) systematically elucidated a collaborative framework of three technological paradigms for machine learning-driven mineral exploration—automated processing, modeling substitution, and data inversion—and revealed their intrinsic correlation mechanisms. The typical case evolution path presented in this study demonstrates that: in the field of intelligent lithology identification based on remote sensing, semi-supervised learning methods based on random forests can effectively handle automatic lithology mapping under the constraint of sparse ground samples, simultaneously generating lithology classification uncertainty quantification maps (Kuhn et al., 2018), thus improving the efficiency of manual interpretation; addressing the bottleneck of geodynamic simulation, the Earth viscoelastic surrogate model constructed by deep neural networks (DeVries et al., 2017) breaks through the limits of traditional numerical simulation, achieving fast, reliable, and high spatiotemporal resolution real-time computation; for sparse data inversion problems, the subspace projection learning framework effectively constrains the non-negative least squares reconstruction solution space through low-dimensional latent space representation (Gupta et al., 2019); at the level of deep geological information analysis, random forests mine features from continuous acoustic data (Rouet-Leduc et al., 2018), transforming microseismic noise into fault boundary failure time identifiers. These advancements collectively mark a paradigm shift in machine learning from an auxiliary tool to a core engine for exploration decision-making.

In this stage, breakthroughs have been achieved in deep feature extraction, significantly enhancing the intelligent capabilities of mineral exploration. For example, in the field of high-resolution remote sensing interpretation, Convolutional Neural Networks (CNNs) have demonstrated significant advantages. Studies by Zhang et al. (2018) on mineralization alteration identification showed that CNN models achieved a recognition accuracy of 92% (compared to 75% for random forest models), with its core advantage stemming from its strong generalization ability to geomorphic heterogeneity. Compared to the limitations of random forests, which rely on artificial feature construction, CNNs achieve adaptive decoupling of alteration information from complex geological backgrounds through hierarchical modeling of local texture features.

In addressing the "high-dimensional, weak anomaly" problem in seismic exploration, unsupervised deep networks such as autoencoders have also made breakthroughs. Valentine and Treampert significantly improved the detection sensitivity of weak anomaly signals by constructing a compact feature representation of seismic waveforms. Current research further combines attention mechanisms with gradient backpropagation visualization techniques (such as Grad-CAM) to maintain model interpretability and lightweight design. It is worth noting that although edge computing combined with lightweight architectures such as MobileNet supports real-time processing on field terminals, data preprocessing still relies on desktop geographic information systems such as ArcGIS, making it difficult to support real-time decision-making in field environments. This bottleneck urgently needs to be overcome through edge-cloud collaborative computing architecture.

Currently, intelligent research in mineral exploration is undergoing a paradigm shift, with its core evolution focusing on two frontiers: breakthroughs in the interpretability of deep learning and the implementation of digital twin technology. Simultaneously, it faces the dual technical challenges of black-box model mechanisms and lightweight deployment. The lack of verifiable causal relationships between deep learning predictions and geological metallogenic theories (Zuo et al., 2024) restricts the geological credibility of exploration decisions; while the real-time processing requirements of multi-source heterogeneous data such as hyperspectral remote sensing and 3D seismic data urgently necessitate the development of domain-adaptive lightweight networks based on Neural Architecture Search (NAS). To overcome these bottlenecks, the academic community is promoting the deep integration of geological knowledge systems and data-driven models: by encoding geological conceptual models such as Zhao Pengda's (2002) "three-linkage" metallogenic prediction theory into topological constraints of graph convolutional networks (Zuo et al., 2022), a cognitive leap from black-box prediction to interpretable gray-box models is achieved; relying on the three-dimensional geological modeling technology developed by Chen Jianping et al. (2024), digital twins are constructed to map the spatial dependence between mineralization response and geological entities in virtual space, forming a verification closed loop under the constraints of physical laws. This knowledge-guided technical path not only significantly improves the geological consistency of intelligent exploration models, but also fosters a new research paradigm of mutual verification and co-evolution between machine intelligence and geological cognition.