AI Learns Geochemistry: New Explainable Model Targets Rare Earth Deposits

Rare earth elements have become the invisible backbone of modern technology, powering everything from wind turbines and electric vehicle motors to smartphones and precision-guided weapons. Yet finding new supplies of these critical metals remains one of the most expensive and uncertain challenges in mineral exploration. A research team led by Kamran Mostafaei of the University of Kurdistan, together with colleagues at Pontificia Universidad Católica de Chile and the University of Southern Queensland, has now unveiled a machine learning framework that promises to change how explorers hunt for one of the most commercially important classes of rare earth deposits. Their approach, described in the journal Earth Science Informatics, is called Explainable Boosted Geochemical Trees, or EBGT, and its central innovation is deceptively simple: it forces the algorithm to obey the laws of geochemistry.

Ion-adsorption rare earth element deposits are among the most sought-after resources in the critical minerals race. Unlike hard-rock rare earth deposits, where the metals are locked within complex mineral lattices that require energy-intensive processing, ion-adsorption deposits host rare earth elements loosely bonded to clay minerals in weathered granite regolith. The metals can be recovered relatively cheaply using simple chemical leaching, which is why these deposits, first recognized as a major resource class in southern China, have become the world’s dominant source of heavy rare earth elements. Weathered granites in Malaysia, formed over tin-bearing S-type granitic provinces, have emerged as promising analogues, and recent studies have documented ion-adsorption-type mineralization across western Peninsular Malaysia.

The problem, however, is that confirming whether a weathered granite actually contains economically adsorbed rare earth elements traditionally requires detailed trace-element analysis. Measuring the full suite of rare earth elements in hundreds of regolith samples demands costly laboratory work, typically involving techniques such as inductively coupled plasma mass spectrometry. In the early stages of exploration, when vast tracts of tropical terrain must be screened quickly, such expense is a serious barrier. The researchers behind EBGT asked a provocative question: could cheap, routinely measured major oxide data, the kind of information available from standard whole-rock geochemistry, be used to reliably flag ion-adsorption rare earth prospectivity before any expensive trace-element assay is performed?

Machine learning would seem like an obvious answer, and algorithms of many kinds have already been applied to mineral prospectivity mapping around the world. But geoscientists have long harbored a deep unease about these tools. Most powerful models, from random forests to gradient boosting machines, operate as black boxes: they produce predictions without revealing why. In a discipline where a wrong drill target can burn millions of dollars, an unexplainable prediction is a hard sell. Worse, unconstrained models can latch onto statistical correlations that have no physical meaning, producing geologically nonsensical rules that happen to fit the training data. The EBGT framework was designed specifically to close this trust gap.

The technical core of EBGT is a gradient-boosted decision tree ensemble, a family of algorithms related to widely used implementations such as XGBoost, in which many weak predictive trees are built sequentially, each one trained to correct the errors of its predecessors. What distinguishes EBGT is that the researchers embedded monotonic constraints directly into the training process. These constraints are hard rules that force the model’s output to move in a specified direction as a given input variable changes, regardless of what the raw data might suggest. The constraints were derived from established geochemical principles of how ion-adsorption deposits form. Because rare earth elements are adsorbed onto clay minerals produced by the chemical weathering of granite, the model was required to treat aluminum oxide, a proxy for clay content, as having a strictly positive relationship with prospectivity. Conversely, silicon oxide, which indicates resistant quartz-rich material that weathers poorly and dilutes the clay fraction, was constrained to have a strictly negative relationship.

Beyond the constraints, the team engineered a set of diagnostic oxide ratios that encode recognizable geological processes into the feature space itself. The ratio of silicon oxide to aluminum oxide serves as a weathering and clay-abundance indicator, tracking how far the parent granite has decomposed into adsorptive clay. The ratio of potassium oxide to sodium oxide acts as a fingerprint of granite type and alteration intensity, reflecting the breakdown of potassium feldspar and the mobility of alkali elements during weathering. The ratio of iron oxide to titanium oxide provides insight into the behavior of accessory phases and the degree of chemical leaching. These ratios were not chosen arbitrarily; each corresponds to a documented control on rare earth adsorption in regolith-hosted systems, meaning that every feature the model sees carries physical meaning.

The framework was tested on major oxide data from weathered granites in Malaysia, a setting where ion-adsorption rare earth potential is actively being evaluated. To establish whether the geochemical constraints cost anything in predictive power, EBGT was benchmarked against four widely used classifiers: logistic regression, random forest, support vector machine, and k-nearest neighbors. The results were striking. EBGT achieved perfect recall, meaning it identified every known prospective sample in the evaluation, alongside an area under the receiver operating characteristic curve of 0.98, a near-perfect discrimination score. Just as importantly, the model showed superior probability calibration, meaning that when it reported, say, an eighty percent likelihood of prospectivity, that number could be taken at face value rather than treated as an opaque score. Calibration matters enormously in exploration, where predicted probabilities are used to rank and prioritize ground.

Accuracy alone, however, was only half the story. To interrogate how the model reached its decisions, the researchers applied SHAP analysis, a technique from explainable artificial intelligence that quantifies each feature’s contribution to individual predictions. The analysis confirmed that the model’s behavior was driven by geochemically meaningful variables rather than spurious correlations. The most influential predictor turned out to be the potassium oxide to sodium oxide ratio, a result consistent with the role of granite composition and feldspar alteration in controlling rare earth adsorption capacity. The constrained relationships for aluminum oxide and silicon oxide held as designed, and the engineered ratios collectively dominated the feature importance rankings. In effect, the model had independently rediscovered, through data, the same geological logic that the researchers had written into its constraints, providing a powerful internal consistency check.

The implications extend well beyond Malaysia. The study demonstrates a general recipe for trustworthy machine learning in the geosciences: instead of asking algorithms to learn everything from scratch, encode what is already known about the system as structural constraints, then let the model discover the remaining patterns from data. This hybrid of domain knowledge and statistical learning produces models that are simultaneously accurate, transparent, and actionable, a combination rarely achieved in mineral prospectivity mapping. For exploration companies, the practical payoff is a cost-effective screening tool that can evaluate weathered granite terrains using inexpensive major oxide assays, reserving expensive trace-element work and drilling for the most promising targets. For the broader critical minerals supply chain, at a moment when governments are scrambling to secure rare earth resources outside a handful of dominant producers, faster and cheaper identification of ion-adsorption deposits could meaningfully reshape where the next generation of mines is found.

The work also speaks to a growing movement within economic geology toward explainable artificial intelligence. As machine learning spreads through prospectivity mapping, noise sensitivity studies, and three-dimensional mineral modeling, researchers increasingly recognize that predictive performance is not enough; models must earn the confidence of the geologists who act on them. By proving that monotonic, geochemically grounded constraints can deliver top-tier performance while guaranteeing physically coherent behavior, the EBGT framework offers a template others can adapt to different deposit types, commodities, and datasets. The black box, in this corner of exploration science at least, has finally been opened, and what lies inside looks remarkably like good geology.

Subject of Research: Explainable machine learning for ion-adsorption rare earth element prospectivity mapping using geochemically constrained models

Article Title: A geochemically constrained explainable machine learning framework (EBGT) for ion-adsorption rare earth elements prospecting

Article References: Mostafaei, K., Kianpour, M. N., Hekmatnejad, A., Jara, J. J., & Shokri, B. J. (2026). A geochemically constrained explainable machine learning framework (EBGT) for ion-adsorption rare earth elements prospecting. Earth Science Informatics, 19(10), Article 175. https://doi.org/10.1007/s12145-026-02225-4

Image Credits: AI Generated

DOI: 10.1007/s12145-026-02225-4

Keywords: rare earth elements, ion-adsorption deposits, machine learning, explainable AI, geochemistry, mineral exploration, gradient boosting, SHAP, weathered granite, Malaysia, prospectivity mapping, critical minerals