AI and DataDriven Decision Making
AI in Mineral Exploration and Geoscience
AI in Mineral Exploration: Separating Hype from Reality Jef Caers, Stanford University, jcaers@stanford.edu Abstract The energy transition has focused global attention on critical mineral exploration, yet the industry has seen declining discovery rates despite increasing investment. This paper argues that the core problem is methodological: the absence of a rigorous, overarching scientific method for exploration. This paper proposes a framework grounded in Bayesian probability and Popperian falsification, enabled by AI, that aims to reduce cognitive bias, limit false positives, and make decision-making more rational and cost-effective. This paper is a summary of “The future of AI in mineral exploration”, which can be found on ArXiv. 1. Introduction Drawing on three decades of experience in data science and decision science applied to Earth resources, I argue that AI in mineral exploration should be seen not as a tool for autonomous discovery, but as an enabler of better scientific method. The paper contrasts the mining sector— where deterministic modeling remains the norm—with oil & gas, where Bayesian uncertainty quantification, ensemble modeling, and AI-assisted decision-making have become standard. Return on investment in exploration has been largely negative for decades. This is not due to a lack of investment or technology, but to the absence of a unifying scientific method that governs how exploration decisions are made. The paper is structured in three parts: (I) a new scientific method, (II) current AI applications and their errors, and (III) the future of AI in mineral exploration. 2. Part I — A New Scientific Method 2.1 Decision-Making in Exploration Exploration is fundamentally an exercise in reducing uncertainty about the existence and economics of ore deposits, operating across scales from continental mapping down to drilling. Decisions are defined as the irrevocable allocation of financial resources. The primary decision at each stage is: what data should we acquire next, and does it justify the cost? This question that is rarely answered with adequate rigor. 2.2 The Problem with Determinism A typical exploration campaign under cover currently proceeds by inverting geophysical data into a single 3D model, interpreting it geologically, and planning drilling based on that single model. The critical flaw is that uncertainty at every step is ignored. The result is a high rate of false positives—boreholes that intersect nothing of interest. When this happens, reasons are rarely analyzed; the project moves on to the next anomaly. 2.3 The Bayesian Approach The Bayesian framework requires explicit quantification of prior uncertainty before data acquisition. Rather than estimating a single ‘true’ model, it demands consideration of all plausible
models, weighted by prior knowledge and updated as data arrives. A common contradiction in practice: practitioners resist the Bayesian prior as ‘too subjective,’ while simultaneously accepting a single deterministic estimate as objective. This logical error is one of the central barriers to progress. A case study (Wei et al., 2025) illustrates the approach applied to the Crystal Lake Gabbro intrusion in Ontario. Multiple geological hypotheses about the geometry of a Ni-Cu intrusion were translated into mathematical models, and hundreds of stochastic realizations were generated—all consistent with existing EM survey data. These revealed that the intrusion could take many plausible shapes, far beyond what a single deterministic inversion would suggest, and fed directly into AI-assisted drilling planning. 2.4 Popperian Falsification Karl Popper’s philosophy holds that scientific theories should be tested by attempting to falsify them, not confirm them. Applied to mineral exploration, this means data acquisition should first aim to falsify geological hypotheses, not merely detect mineralization. Bayes and Popper can be combined into a ‘Popper-Bayes’ protocol: • Domain geologists generate multiple conceptual model hypotheses from existing data. • Numerical modelers translate these into mathematical representations with quantified uncertainty. • Prior distributions are tested against observed data through falsification (statistical tests). • Bayesian inversion integrates surviving hypotheses with new data. • Decisions on further data acquisition are made from the resulting uncertainty quantification. Executing this protocol requires all domain experts—geologists, geophysicists, geochemists, geostatisticians, and decision scientists—to work in an integrated workflow, something that does not reflect how exploration companies are currently organized. 3. Part II — Current AI Applications and Common Errors 3.1 Predictive AI: The Data Curation Problem The most critical challenge in applying machine learning to mineral exploration is not algorithm selection—it is dataset curation. Three pervasive errors that generate false positives: • Scale mismatch: core samples (centimeter-scale) are compared with geophysical data (tens to hundreds of meters), artificially inflating apparent correlations. • Deterministic interpolation: interpolating geophysical and geochemical data deterministically reduces variance, smoothing out variability that is essential for accurate predictions. Stochastic simulation should be used instead. • Unjustified data pooling: merging data from geologically distinct zones produces spurious correlations. Models trained in one area cannot be transferred to another without explicit testing. A prospectivity mapping case study (Wang et al., 2020) quantifies the difference: deterministic interpolation yielded a prediction recall of 0.4 against known deposits; stochastic simulation raised it to 0.7. Stochastic outputs also enable risk-return analysis—identifying zones with high expected value alongside manageable uncertainty.
3.2 Decision-Making AI: Rare but Impactful While predictive AI is increasingly applied, decision-making AI remains rare in exploration. The most significant real-world example is the characterization of the Mingomba copper deposit in Zambia (Dempsey, 2024), where a Partially Observable Markov Decision Process (POMDP) was used to sequentially plan borehole campaigns. The explicit objective was to first falsify geological hypotheses about high-grade mineralization geometry before targeting grade and tonnage. This proved substantially more efficient than conventional grid-based drilling (Mern and Caers, 2023). The rarity of such applications is attributed to the fact that decision-making requires irrevocable resource commitments, something outside the comfort zone of academic research—and to the short-term incentive structures that dominate junior mining finance. 4. Part III — The Future of AI in Mineral Exploration 4.1 Human-in-the-Loop Data Science AI will be that of augmentation, not replacement. AI should remove tasks that humans perform poorly and amplify what they do well. Humans excel at generating process-based conceptual hypotheses, interpreting spatial relationships between geological units, and building geologically plausible models. They perform poorly at synthesizing large document sets, discovering patterns in high-dimensional data, avoiding cognitive bias, and quantifying uncertainty rigorously. The ideal system is one where AI continuously interacts with domain experts to generate, test, and update multiple competing geological hypotheses, outputting results in a decision-ready format. High-priority AI tools include: stochastic multi-physics inversion, generative AI for realistic 3D geological modeling, high-dimensional anomaly detection, surrogate models to accelerate forward simulations, and data science methods designed around falsification rather than data fitting. 4.2 Optimal Sequential Data Acquisition The Popper-Bayes framework provides a principled basis for deciding what data to acquire next. Two quantitative metrics can be used for ranking acquisition campaigns: • Value of Information (VOI): ranks surveys by expected economic return. • Efficacy of Information (EOI): a dollar-free metric quantifying how much a proposed survey design will reduce uncertainty about a target quantity, on average, across all possible data outcomes. When combined with POMDP frameworks solved by Monte Carlo Tree Search, these enable fully automated sequential drilling plan optimization. The proposed hierarchy for sequential objectives is: first drill to falsify geological hypotheses, then to constrain deposit volume, then to estimate grade. 4.3 Barriers to Implementation Four structural barriers that currently prevent the framework from being adopted at scale: • Organizational structure: major companies are organized by discipline rather than by decision problem, with data scientists making up roughly 1% of staff. Cross-disciplinary, decision-focused teams are needed.
• Deterministic culture: single models become entrenched as institutional truth over time. Uncertainty is culturally equated with incompetence rather than scientific honesty. • Software ecosystem: current commercial software reinforces deterministic workflows. Highperformance computing (needed for Monte Carlo sampling) is expensive and inaccessible to small companies. • Junior mining model: most exploration is funded on a short-term, one-borehole-at-a-time basis. This structurally discourages the multi-hypothesis, long-horizon approach the paper advocates. A portfolio-based investment vehicles can serve as a structural fix—analogous to diversified equity portfolios—where risk and return are traded off across a set of prospects rather than evaluated one at a time. Extending the JORC reporting standard to require uncertainty quantification, modeled on oil & gas practice, is much needed. 5. Sustainability and Education 5.1 Sustainable Exploration Environmental and social sustainability should be integrated into exploration decision-making from the outset, not appended after resource delineation. The key observation is that sustainability impact is inversely proportional to deposit grade: a high-grade underground operation has a far smaller footprint than a low-grade open pit. This means that targeting high-grade deposits is simultaneously the most economical and most environmentally responsible strategy. Sustainability metrics—social license, jurisdictional regulations, environmental impact—should be quantified alongside economic metrics in portfolio-level analysis. 5.2 Knowledge and Education The arrival of large language models has accelerated a long-overdue transformation in how domain knowledge is created, accessed, and taught. Of concern is the declining enrollment in economic geology and mining engineering in Western countries, the risk of losing century-old knowledge stored in aging reports, and the absence of integrated ‘Mineral Exploration’ curricula that combine geoscience with decision science and data science. Priority recommendations include digitizing and open-sourcing national geological databases, creating web-accessible platforms for historical exploration data, and redesigning curricula around problem-solving rather than discipline-only training. Building a shared language between geoscientists and data scientists is identified as essential. 6. Summary The exploration industry’s declining discovery rate is primarily a methodological problem, not a resource scarcity problem. The proposed fix—Popper-Bayes exploration enabled by AI—replaces single-model determinism with multi-hypothesis uncertainty quantification, and replaces ad hoc drilling with sequentially optimal data acquisition. Several components of this framework already exist in research-grade implementations. The bottleneck is integration: assembling them into a usable platform, combined with cultural change in the industry, new financial instruments, and updated regulatory standards. This change will not happen overnight, and that both the industrial and academic communities must move together for it to take hold.
Caers, J., 2025. The future of AI in critical mineral exploration. arXiv preprint arXiv:2512.02879.
AI-Enabled Mineral Exploration: Toward Faster, Cheaper, and Smarter Discovery *J. Mern PhD1 1Terra AI, CA, USA (*Presenting author: jmern@TerraAI.Earth) Abstract Exploration decisions are made under severe geological uncertainty, sparse data, and long feedback cycles, yet early-stage outcomes disproportionately influence project cost, risk, and development timelines. This paper evaluates how modern AI methods can be applied to improve exploration decision-making by integrating heterogeneous data, explicitly modeling subsurface uncertainty, and optimizing sequential exploration actions. We present a structured assessment of AI applications in exploration, focusing on three core capabilities: multi-modal data fusion, probabilistic uncertainty modeling, and closed-loop decision optimization. We discuss methodological considerations related to transparency, interpretability, and validation that are necessary for responsible deployment in operational settings. To quantify impact, we apply these methods to a structurally complex copper porphyry system using legacy drilling, geophysical, and geological data. A neural-network-based generative modeling framework is used to produce probabilistic three-dimensional subsurface realizations and associated resource distributions, which are then coupled with stochastic drilling optimization to guide sequential targeting. The case study demonstrates that AI-derived resource models achieve substantially tighter uncertainty bounds than conventional expert-driven range analyses while maintaining geological consistency with available data. Sequential AI-guided drilling reaches resource models statistically comparable to those obtained through conventional targeting using approximately 45% of the drilling meters. We further show that improvements in early-stage resource precision can significantly reduce false-positive advancement decisions and that accelerated resource upgrading could materially affect projected copper supply trajectories. KEYWORDS World Mining Congress, Mineral exploration; probabilistic modeling; artificial intelligence; data integration; copper; uncertainty quantification 1. INTRODUCTION The term “AI”, an acronym for Artificial Intelligence, has recently become a focus of minerals, mining, and critical resource development across the world. Many see AI as having the potential to solve long-standing problems in the industry, ranging from exploration through production. Like in most other emerging fields of application, there is not yet a consensus understanding of what
the term AI specifically means, what methods it refers to, and generally what its capabilities (and limitations) are. Developing a deeper understanding of these questions is critical for leaders in the minerals world to be able to meaningfully assess which new AI technologies to adopt, and to build the trust and verification methods necessary to deploy them correctly in their operations. In this paper, we focus on AI applications in exploration. Exploration is inherently difficult due to extreme data sparsity, high geological uncertainty, non-unique interpretations of indirect geophysical measurements, and long feedback cycles between decisions and outcomes. Critical decisions are often made early, when information is incomplete and heterogeneous, yet these decisions disproportionately determine project cost, risk, and ultimate success. These characteristics make exploration a natural application area for AI systems that can systematically integrate diverse data types, quantify uncertainty, and support consistent decision-making under uncertainty. 2. AI METHODS IN EXPLORATION AI has been applied across a wide range of exploration tasks, spanning sample-scale geochemical analysis through to basin- and district-scale structural interpretation. Within industry practice, adoption has largely concentrated in two application areas. The first is discovery identification, where AI methods are applied primarily in 2D maps over large spatial extents to identify prospective regions for follow-up exploration and testing (e.g., VRIFY’s AI-assisted discovery workflows; VRIFY, 2025). While promising, these approaches face long validation timelines due to the inherently stochastic and sparse nature of mineral discovery — even substantial improvements in targeting efficiency can still correspond to low discovery ratios in practice — and because regional-scale applications must contend with strongly non-stationary relationships between geophysical responses and underlying geology that challenge even modern supervised learning methods. The second area is exploration and modeling, which focuses on understanding mineral systems at the prospect- to camp-scale, typically in three dimensions and with richer, sitespecific datasets that help overcome some transfer limitations of broad 2D targeting. This domain remains more nascent and generally requires highly customized workflows rather than off-theshelf models, but offers greater potential to integrate uncertainty-aware modeling and geological constraints (Caers, 2025; Mineral-X, 2025). We explore this latter area in greater detail in the following sections. 2.1 Key Capabilities in Exploration Across exploration and subsurface modeling workflows, three underlying AI functions have emerged as particularly impactful: data fusion, uncertainty modeling, and decision optimization. Together, these capabilities address key structural limitations of conventional exploration workflows and enable more systematic reasoning under uncertainty. Data fusion refers to the joint analysis of multiple heterogeneous data sources—such as geophysics, geochemistry, geological mapping, and drilling data—within a single modeling framework. A central limitation of traditional exploration practice is that these data types are typically analyzed using separate, loosely coupled quantitative processes, often requiring subjective reconciliation at later stages. AI-based data fusion methods enable the simultaneous analysis of these data, allowing models to capture higher-order interactions and non-linear relationships that are difficult to represent using manual or sequential approaches. In mineral exploration, fused data representations have been successfully applied to regional prospectivity
mapping (Carranza & Laborte, 2015; Cracknell & Reading, 2014) as well as to more direct threedimensional geological modeling and uncertainty analysis at the prospect and camp scales (Jessell et al., 2018). Uncertainty modeling addresses the central objective of exploration: reducing uncertainty about the subsurface in order to support better decisions. To enable rigorous optimization and comparison of alternative exploration strategies, uncertainty must be explicitly represented rather than treated implicitly or qualitatively. Modern AI approaches to uncertainty modeling—often grouped under probabilistic modeling—include methods such as normalizing flows, deep generative models, variational latent variable models, and Bayesian neural networks (Rezende & Mohamed, 2015; Kingma & Welling, 2014; Neal, 2012). These methods are designed to operate in highdimensional, uncertain domains and provide not only predictions but also calibrated measures of confidence. When applied to subsurface problems, they enable the generation of three-dimensional uncertainty maps and ensembles of plausible geological realizations, supporting both interpretation and downstream decision-making. Decision support and optimization build on fused data representations and probabilistic models to guide exploration actions under uncertainty. In many high-stakes domains—such as robotics and autonomous systems—AI models are explicitly embedded within closed-loop decision frameworks that combine learning, simulation, and active planning. Methods such as reinforcement learning and Monte Carlo tree search are used to evaluate large sets of possible actions and future states, optimizing long-term objectives rather than single-step criteria (Sutton & Barto, 2018; Browne et al., 2012). In exploration, analogous decision problems arise when selecting drill targets, survey designs, or sequencing strategies under uncertain geology. Formal results in Bayesian decision theory demonstrate that closed-loop, adaptive strategies can outperform open-loop or one-shot approaches—such as static information-theoretic targeting— both in expected value and regret bounds (Frazier, Powell, & Dayanik, 2008). As a result, AIenabled decision optimization represents a critical step toward scalable, repeatable exploration decision-making. 2.2 Trust and Validation For AI systems to be adopted responsibly in exploration, trust and interpretability are essential. Exploration decisions are capital-intensive and therefore require not only predictive performance but also clear understanding how AI outputs are generated. Without appropriate transparency and validation, even technically strong AI models risk being underutilized or misapplied. A foundational requirement for trust is clarity about the underlying methodology. AI providers, in most cases, should be able to explain what class of methods is being used, what classes they are training, and how those methods differ from established techniques beyond. In practice, AI developers should be able to share a fair amount of clarifying detail without unreasonable IP risk. This transparency can help distinguish AI approaches from the rebranding of long-established data science or geostatistical methods—such as kriging or principal component analysis—as “AI.” At the same time, it is important to recognize that some degree of black-box behavior is unavoidable in modern AI systems. State-of-the-art neural models commonly contain hundreds of millions to billions of parameters and rely on deeply non-linear transformations. Unlike traditional
geophysical forward models, these systems do not admit closed-form solutions or explicit governing equations. As a result, full mechanistic interpretability should not be expected. Input– output intuition, however, should still be preserved. AI predictions should be retrospectively interrogable in a manner that is consistent with geological reasoning. When a model extrapolates beyond conventionally-constrained data, there should be a plausible geological trend, structural control, or data-derived signal that informs that extrapolation. Expert review of AI outputs should be able to identify whether predictions are driven by meaningful data relationships or by spurious correlations. Finally, explicit validation procedures are critical. Users should understand how AI models were validated prior to deployment. While no single validation approach is sufficient in isolation, robust AI workflows typically employ multiple checks. At a minimum, models should demonstrate appropriate fit to existing data, ideally through hold-out or blind testing where the model is evaluated against observations that were not used during training or calibration. In cases where direct data interrogation is not possible—such as in regional prospectivity mapping—additional scrutiny is warranted, including sensitivity analyses, consistency checks across data subsets, and comparison against independent geological interpretations. Deploying an unconverged or overfit model is an easy mistake that can lead to extremely poor outcomes. 3. EXPLORATION CASE STUDY In this section, we present the results of a case study conducted by Terra AI on data collected by a conventional exploration program. The visuals and numerical results are presented for an analogous synthetic case that retain their relative magnitudes for comparison purposes. 3.1 Case Study Setup The case study focuses on a structurally complex copper porphyry system characterized by faulting and post-emplacement structural modification. The primary objective was to characterize ore-body geometry, grade, and tonnage with sufficient confidence to support an investment decision. The analysis incorporated a heterogeneous set of exploration data. Inputs included legacy drilling data from adjacent areas, comprising lithological logs, wireline measurements, and geochemical assay data. These were complemented by multiple geophysical datasets—including gravity, magnetotellurics (MT), direct current induced polarization (DCIP), and transient electromagnetics (TEM)—as well as surface lithological maps, topography, and relevant regional geological literature. AI-based modeling was applied through a neural-network-driven data fusion and generative modeling framework, producing probabilistic ensembles of three-dimensional subsurface models. These ensembles were used to generate resource estimates across multiple statistical quantiles, explicitly representing uncertainty in geometry, grade distribution, and total contained copper. A sequential, stochastic drilling optimization approach was then employed to generate drill targets designed to maximize expected uncertainty reduction in total contained copper above a specified economic cutoff.
Performance was evaluated through comparison with conventional exploration workflows. First, AI-driven resource estimates were compared against traditional resource modeling results to assess relative precision, contrasting objective, probabilistic estimates with expert-driven assessments produced independently by different practitioners. Second, drilling performance was evaluated by measuring the amount of drilling required for the AI-optimized strategy to reach comparable levels of precision and accuracy to the resource model derived from the full set of conventionally targeted drilling. This dual comparison enabled assessment of both modeling effectiveness and exploration efficiency. 3.3 Results First, we compared the efficacy of the multi-modal generative modeling by comparing the resulting resource estimate to an expert-developed resource estimate using the same data. Figure 1 shows the resource distribution generated by the AI system using the available legacy data, prior to any additional drilling on the target. The histogram was constructed by computing total contained copper for each geological realization in an ensemble of approximately 1,000 probabilistic subsurface models. All realizations achieved approximately 85% agreement with the input data and exhibited geologically plausible variability away from data-constrained regions. For comparison, the conventional uncertainty assessment is shown below the histogram as a boxand-whisker plot, with box quantiles defined at P30 and P70 and whiskers extending to the minimum and maximum case values. While both representations incorporate the same expert geological input, the histogram encodes substantially more information by explicitly weighting each realization according to data fit within the AI-driven framework. As a result, the P30–P70 spread of the AI-derived resource distribution is approximately three times narrower than that of the conventional expert-based estimate. Figure 1 - Comparison of legacy-data derived resource distribution to expert-specified analysis range. We next assessed the performance of AI-guided drilling optimization relative to conventionally targeted drilling. The AI system generated drill targets in batches of four, consistent with the operational constraints of a four-rig drilling program. Drilling under the AI-guided strategy was terminated once the resulting resource model was determined to be statistically comparable to the
resource model derived from the full set of conventionally targeted drill holes. This condition was met after the AI-guided program had targeted drilling meters equivalent to approximately 45% of the total footage drilled in the conventional campaign. Figure 2 compares the resource models generated from the original legacy dataset, the AI-guided drilling campaign, and the conventional drilling campaign. The resulting P10, P50, and P90 ore shells exhibit a high degree of consistency across all three models, with differences in total lowgrade volume of approximately 6%. Figure 3 presents the corresponding resource distributions, indicating that the AI-derived model reproduces the central tendencies and uncertainty bounds of the conventional model within reasonable tolerances. Taken together, these results suggest that AI-guided drilling optimization can achieve comparable levels of resource characterization while substantially reducing drilling requirements. At the project scale, this reduction translates into meaningful decreases in capital expenditure, development timelines, and exposure to exploration risk, with potential implications for both project economics and portfolio-level capital efficiency. Figure 2 - Low and high grade resource ore-shells for the P30, P50, and P70 cases of the predrilling model, AI-targeted drilling model, and conventionally targeted drilling model.
Figure 3 - Resource distributions for the pre-drilling model, AI-targeted drilling model, and conventionally targeted drilling model. CONCLUSIONS AND FUTURE WORK This paper examined the application of AI to mineral exploration, with a focus on data fusion, uncertainty modeling, and decision optimization, and demonstrated their impact through a largescale copper porphyry case study. The results show that AI-driven, probabilistic subsurface modeling can produce materially more precise resource estimates than conventional expert-based uncertainty analyses, while maintaining geological plausibility and consistency with available data. In addition, sequential, AI-guided drilling optimization achieved resource models statistically comparable to those derived from conventional drilling campaigns using approximately 45% of the drilling meters, indicating substantial potential for accelerating resource characterization and reducing exploration cost and risk. At the system level, these improvements translate into meaningful implications for project economics, development timelines, and long-term copper supply constraints. Despite these encouraging results, AI in exploration remains at a relatively early stage of maturity. Subsurface systems are complex, data are sparse and heterogeneous, and outcomes are strongly conditioned by geological context. As a result, realizing the full potential of AI will require continued bi-directional learning between the AI research community and the resource sector. Advances in model architecture, uncertainty quantification, and decision frameworks must be informed by geological realism and operational constraints, while domain practices must evolve to effectively integrate probabilistic outputs and adaptive decision-making into established workflows. Future work spans several important directions. At regional scales, two-dimensional prospectivity mapping remains an active area of research, with opportunities to improve robustness,
generalization across geological settings, and validation methodologies. At the project scale, significant potential exists to extract additional information from geophysical data by applying AI directly to raw or minimally processed measurements, rather than relying primarily on derived or interpreted products as inputs. More broadly, extending AI-driven frameworks to additional exploration problems—such as survey design, adaptive data acquisition, and integrated economic decision-making—represents a promising avenue for further impact. Continued progress across these fronts will be critical to transitioning AI from promising pilot applications to a scalable, trusted component of modern exploration practice. ACKNOWLEDGEMENTS The author would like to thank the complete team at Terra AI, advisors and past collaborators at Stanford University, and our entire set of industry partners for providing insights and experiences inspiring this work. REFERENCES Caers, J. (2025). The future of AI in critical mineral exploration. arXiv. https://arxiv.org/pdf/2512.02879.pdf Mineral-X. (2025). Scientific research. Mineral-X. https://mineralx.stanford.edu/scientificresearch VRIFY. (2025). Mineral exploration tech report 2025. VRIFY. https://vrify.com/resources/mineral-exploration-tech-report Browne, C. B., Powley, E., Whitehouse, D., Lucas, S. M., Cowling, P. I., Rohlfshagen, P., … Colton, S. (2012). A survey of Monte Carlo tree search methods. IEEE Transactions on Computational Intelligence and AI in Games, 4(1), 1–43. https://doi.org/10.1109/TCIAIG.2012.2186810 Carranza, E. J. M., & Laborte, A. G. (2015). Data-driven predictive mapping of mineral prospectivity using machine learning methods. Ore Geology Reviews, 71, 804–818. https://doi.org/10.1016/j.oregeorev.2014.11.014 Cracknell, M. J., & Reading, A. M. (2014). Geological mapping using remote sensing data: A comparison of five machine learning algorithms. Computers & Geosciences, 63, 22–33. https://doi.org/10.1016/j.cageo.2013.10.008 Frazier, P., Powell, W., & Dayanik, S. (2008). A knowledge-gradient policy for sequential information collection. SIAM Journal on Control and Optimization, 47(5), 2410–2439. https://doi.org/10.1137/070693424
Jessell, M. W., Aillères, L., de Kemp, E., Lindsay, M., Wellmann, F., Hillier, M., & Laurent, G. (2018). Next generation 3D geological modeling and inversion. Geological Society, London, Special Publications, 451(1), 261–272. https://doi.org/10.1144/SP451.7 Kingma, D. P., & Welling, M. (2014). Auto-encoding variational Bayes. Proceedings of the International Conference on Learning Representations (ICLR). Neal, R. M. (2012). Bayesian learning for neural networks. Springer Lecture Notes in Statistics. https://doi.org/10.1007/978-1-4612-0745-0 Rezende, D. J., & Mohamed, S. (2015). Variational inference with normalizing flows. Proceedings of the 32nd International Conference on Machine Learning (ICML), 1530–1538. Sutton, R. S., & Barto, A. G. (2018). Reinforcement learning: An introduction (2nd ed.). MIT Press. International Energy Agency. (2025). Copper – Analysis. IEA. https://www.iea.org/reports/copper S&P Global Energy. (2025). Copper in the age of AI: The challenges of electrification. S&P Global. https://www.spglobal.com/en/research-insights/special-reports/copper-in-the-age-of-ai
IDENTIFICATION OF TARGETS IN A SILVER VEIN DISTRICT USING MACHINE LEARNING *J.B. Harris¹, L. Lugones¹, A. Rochefort¹, T. Rodriguez¹, M. Contreras¹, J. Muñóz¹ ¹Mineral Forecast, Santiago, Chile. (*Presenting author: jacqueline.harris@mineralforecast.cl) ABSTRACT Mineral exploration in mature brownfield districts faces the critical challenge of identifying new high-potential targets amidst complex geological settings where traditional methods often reach their limits, leading to increased costs and slower discovery rates. This study addresses this problem by implementing an advanced Machine Learning (ML) framework to integrate and analyze multi-dimensional data from a low-sulfidation epithermal silver district in Mexico. The solution involved a two-stage methodology: first, a technological validation using clustering to verify the model’s predictive capabilities in known sectors (Veta Pérez), and second, the evaluation of 50 different ML architectures to select the optimal K-Neighbors model. This specific architecture was chosen for its superior ability to maximize precision and minimize false positives, which is crucial for high-cost drilling decisions. The model successfully integrated 45 variables, including geophysics, lithology, and structural data from over 230,600 meters of historical drilling, to identify 30 new areas of interest within a 248 km² area. The practical impact of this approach is demonstrated by a 94.74% accuracy and a 70.86% precision rate in the final predictions. Most significantly, out of 34 targets generated by the model and subsequently drilled, 20 yielded positive mineralization (Ag > 50 ppm), achieving a 59% success rate. These results prove that the proposed ML workflow provides a reliable, data-driven guide that significantly optimizes exploration budgets and timelines, directly contributing to the industry's goal of delivering minerals faster and smarter. This methodology directly addresses the global challenge of delivering minerals faster and smarter by transforming historical data into high-confidence predictive model exploration targets. KEYWORDS Machine Learning, Mineral Exploration, Silver Veins, Predictive Modeling, WMC 2026. 1. INTRODUCTION We identify Several preliminary tests were conducted new predictive model exploration targets through the synthesis of multi-dimensional data is a primary driver for integrating Machine Learning (ML) into the mining industry. This project is situated near an active underground operation in Mexico with a history dating back to 1757. The deposit is a low-sulfidation epithermal vein system characterized by rigorous structural control. The primary objective is to identify new areas of interest across 248 km². The challenge lies in integrating 45 disparate variables including lithology, hydrothermal alteration, and
Figure 1A.- Analysis of rock types and structural orientations showing favorable correlation with high-grade silver mineralization. 1B.- Orientation of structures favorable for mineralization, given their occurrence in the vicinity as the silver grade increases. geophysics to establish a framework that enhances target identification in rugged terrain. 2. AVAILABLE INFORMATION A total of 45 variables were integrated. The dependent variable is silver (Ag) concentration with a threshold of 50 ppm Ag. 2.1 Dependent Variable The dependent variable defined for this study is silver (Ag) concentration, sourced from the geochemical database of all available drill holes. A threshold of 50 ppm Ag was established as the cut-off grade for the predictive exercise. This specific concentration represents the minimum economic grade acceptable for the processing plant and was defined as the primary target by the mine’s exploration team 2.2 Exploratory Variable The predictive or explanatory variables integrated into the model include lithology (both surface and subsurface), hydrothermal alteration, structural data, and geophysics (magnetometry and radiometry). The effective projection area is constrained by the overlap of these datasets; in this instance, the combined explanatory variables cover a surface of 109 km² out of the total 248 km² of the district. 2.2.1 Lithology Comprises 17 surface and eight subsurface units. Units such as ‘kpa’ and ‘tcr’ show favorable correlation with mineralization Figure 1A. 2.2.2 Alteration Eight classes derived from ASTER satellite imagery, including argillic, phyllic, and silica categories. 2.2.3 Structures Categorized by strike orientation. The district features five structural blocks where vein orientations shift from NE in the west to NS in the east Figure 1B. 1A 1B
2.2.4 Geophysics Includes 13 layers of magnetometry and radiometry, processed with a 30-meter grid cell size 3. METHODOLOGY The process followed two stages: 3.1. First Stage Clustering the area into 5 groups (77,198 samples). The Veta Pérez sector (22.10% of data) withheld for validation, while the remaining 77.89% was used for training. 3.1.1. Clustering The amount of available drill hole data was considered sufficient to perform clustering rather than random selection. The team conducted several preliminary tests were conducted before arriving at the final configuration. We clusters the area into five clusters, corresponding to 77,198 drill hole samples. Clusters 0, 1, 2, 3, and 4 represent 14.73%, 22.10%, 27.15%, 5.03%, and 30.99% of the total data, respectively Figure 2A. This segmentation is critical as it ensures that the model’s predictive logic is grounded in the district's specific geological domains rather than being a generalized mathematical abstraction. 3.1.2. Test Area Selection After testing several areas, the Veta Pérez sector proved to be the best choice and was selected as the definitive test area, containing 22.10% of the available data. These samples were withheld from the database Figure 2B. By withholding a well-known sector like Veta Pérez, we create a rigorous 'blind test' that proves the model can accurately identify mineralization in areas it has not yet processed. 3.1.3. Training Area The training area comprises 77.89% of the available samples and corresponds to clusters 0, 2, 3, and 4. This data was used to learn from the mineralization phenomenon to be explained, and this learning was then applied to make predictions in the Veta Pérez area, where the information had been previously withheld. 3.1.4. First Stage Evaluation The validation of the different architectures includes both a visual and a numerical component. The visual part aims to identify how well the model can recognize known zones. This is done through a point-by-point comparison between the predicted sample and the actual drill hole sample. The numerical part seeks to identify, through various parameters, which of the applied architecture best responds to the identification of the mineralization phenomenon. The metrics used are precision, F1-score, recall, accuracy, and specificity. Establishing these baseline metrics
(Precision, Recall, Accuracy) allows for a transparent assessment of risk before deploying the model for expensive greenfield drilling campaigns. 3.2. Second Stage Architecture selection. We executed 65 iterations were performed, evaluating 50 different architectures per iteration to find the optimal model using 100% of the data. 3.2.1. Machine Learning Architecture In the second, ML architecture was generated using all available information. A total of 65 different iterations were performed to arrive at an optimal architecture. These iterations included changes in variables and in the layers used. Additionally, each iteration involved a battery of 50 different ML architectures, which were ranked, and the top 5 performers were subsequently visualized. From the set of architecture identified in the first stage, the optimal one was selected and used to make predictions across the entire area. The metrics used were precision, recall, F1score, accuracy, and specificity. Evaluating 50 different architectures per iteration ensures that the final ML Architecture K-Neighbors selection is the most robust and stable tool for handling the specific noise and complexity of the district's geophysical and lithological data. 4. RESULTS AND DISCUSSION The model successfully predicted the mineralization of the targets. The selected ML Architecture was K-Neighbors, due to its ability to prioritize accuracy and minimize false positives. The mineralized lithological layers are present in all five blocks identified in the district. The study made it possible to identify new mineralized bodies, corroborate existing ones, and generate strategic knowledge for future exploration planning. The results of the first stage were successful, satisfactorily recognizing the Veta Pérez zone (Figure 3A). The resulting metrics were: 42.27% precision, 1.83% recall, 3.50% F1-score, 86.76% accuracy, and 99.62% specificity. The selected architecture was the one that maximized precision, prioritizing the reduction of false positives. After evaluating 50 different ML Architecture, KNeighbors stood out as the option that best met this criterion, offering an efficient solution aligned with the analysis objectives. 2A 2B
Figure 3A ML target architecture results for the entire Ag district, highlighting newly identified areas of interest and successful drilling validations. 3B Details of one of the district targets.in 3D. The ML architecture generated in the second stage, with 100% of the data, was able to identify several zones with high mineralization potential. To minimize the presence of false positives, the focus was on maximizing precision. A total of 34 areas of interest generated through ML have been drilled so far, of which 20 have yielded positive results (with Ag concentrations exceeding 50 ppm, in some cases over 200 ppm). This equates to a drilling success rate of 59%. Additionally, the false positive zones drilled (the ML model indicated no mineralization, but there was anyway) are valuable inputs to keep tuning the ML model and improve its results. We employed explainability tools to understand how ML architecture learned. Geophysical variables contribute the most to the model, followed by lithological variables, and then structural variables. The resulting metrics for the second stage are: 70.86% precision, 28.29% recall, 40.43% F1-score, 94.74% accuracy, and 99.22% specificity. In practical exploration terms, the 70.86% precision indicates that 7 out of 10 targets predicted by the model are likely to be mineralized, which significantly optimizes the allocation of drilling budgets by reducing 'dry' holes. The project's results supported the exploration of new veins, and the global understanding of the variables provided value for understanding the emplacement of the deposit's mineralization. From an exploration perspective, prioritizing precision over recall means that while the model may not find every vein in the district, the targets it does identify have a high probability of success (59%), significantly reducing the financial risk of 'dry' drill holes. 4. CONCLUSIONS Despite being an area with a long history of work, the epithermal system still contains zones with vein-type structures that have not been sufficiently explored and could extend the life of mine. The occurrence of these structures has been identified both along strike and at depth. The study is limited to the extent of the areas covered by the data layers used. An opportunity for improvement involves extending existing studies to areas without coverage, such as the structural information layer, which yielded very good results during the work process. With this new information, the model can be re-run to extend the results to new areas of interest and adjust the outcomes based on the new data provided to the model. 3A 3B
There are at least two preferential orientations for the emplacement of mineralized structures within the district (ENE and NE). The study enhanced the understanding of the data through the generation of graphical analyses of the variables. Additionally, it improved the comprehension of how to define new zones through the generation of explainability graphs, providing valuable information to the exploration team for extending studies in the district. These new knowledge parameters allow for more informed decisions about which tools to extend or implement in the future, thereby maximizing the use of budgets. The study successfully identified extensions of known mineralized areas and new zones to explore, consolidating the application of the methodology and technology as an additional layer of passive information that can be used strategically for defining future areas of interest. The primary metric used to select the best ML architecture was precision, over recall, F1-score, accuracy, and specificity. This was done with the objective of minimizing the presence of false positives in the models. A key aspect to highlight in this work is that, despite the relatively low recall and precision, the fundamental goal in the context of mineral exploration is the ability to identify mineralized bodies. In this case, it was not necessary to achieve millimeter-level precision, as small displacements of a few meters in the estimated location do not affect the practical value of the model. The important thing was to provide a reliable guide to focus efforts on areas with higher potential, thus optimizing resources and time in the early stages of the predictive model exploration process Figure 3B. ACKNOWLEDGEMENTS The authors thank First Majestic for permission to present this case study and the technical team for their support in developing these predictive models. REFERENCES Albrecht, T., González, I., & Klump, J. (2021). Using Machine Learning to Map Western Australian Landscapes for Mineral Exploration. MDPI, v. 10. Mendoza, R., Merino, R., Vasquez, M., & Rosario, P. (2020). NI 43-101 Technical Report, San Dimas Silver/Gold Mine, Durango and Sinaloa States, Mexico.
Identification and measurement of minerals on thin sections images of intrusive igneous rocks using Deep Learning *S.D. Paucar1 1Department of Geology, Central University of Ecuador, Ecuador, (*Presenting author: stalynpaucar271828@gmail.com) ABSTRACT Artificial Intelligence (AI) has moved beyond theory and is now applied in everyday contexts such as politics, economics, medicine, social networks, web browsers, and academic research. Since the early 21st century, Deep Learning has also been increasingly applied within the Earth Sciences. In this research, the Detectron2-v0.6 model was used to identify eight minerals on intrusive igneous rock thin sections, while the scikit-image library was employed to obtain quantitative measurements of the identified minerals. The methodology follows a standard AI model workflow. A total of 400 images were collected from the British Geological Survey and The Open University databases, and 62178 minerals were labeled in the Roboflow platform. These data were used to create four datasets for training and validation, along with an external evaluation dataset of 20 images. All experiments were conducted using Google Colab. The best performance was achieved using the XPL-10X dataset, with a total loss of 1.15 and an average precision (AP, IoU = 0.5-0.95) of 15.7%. Scikit-image showed relative errors between 0-3%, suitable for quantitative analysis. An interactive application was developed in Google Colab and Hugging Face-Streamlit, allowing users to upload images and generate automatic PDF reports with mineral segmentation, identification, mineralogical analysis, and crystal size distribution histograms. KEYWORDS Minerals, Intrusive igneous rocks, Deep Learning, Convolutional Neural Network, Detectron2 1. CONTEXT AND PROBLEM STATEMENT In the last decade, the AI model with the greatest impact has been the Convolutional Neural Network (CNN), whose rise is mainly due to the use of Graphics Processing Units (GPUs) (Bhatnagar, 2023). Koeshidayatullah et al. (2020), using the VGG16 and InceptionResNetV2 models, automated the petrography of carbonate rocks. In sandstones, Liu et al. (2022) employed the ResNet, DenseNet, Mask-RCNN, Generative Adversarial Network (GAN), and Path Aggregation Network (PANet) models to identify minerals, analyze porosity, weathering, sorting, roundness, and grain contacts. Currently, three open-source models represent the state of the art in object segmentation and/or identification: Detectron2, SAM, and YOLO. Sitar & Leary (2023) use Detectron2 to segment zircons, while Zaki et al. (2023) applied it to the segmentation of alite and belite in clinker. Azzam et al. (2024) developed GrainSight, based on SAM, for grain segmentation in sandstones, and the company DiUS, together with Solve Geosolutions, implemented Datarock to analyze drill
RkJQdWJsaXNoZXIy MTM0Mzk2