New Framework Maps Hidden Mineral Deposits While Quantifying Data Uncertainty

Locating ore bodies buried deep beneath the surface has long been a high-stakes guessing game, but a study published in Natural Resources Research aims to make the process more rigorous. Researchers have unveiled a data-uncertainty-aware and explainable framework for three-dimensional mineral prospectivity mapping, a technique that integrates multi-source geological and geophysical data to delineate subsurface exploration targets. The work addresses a persistent blind spot in the field: while 3D mapping models have grown increasingly sophisticated, most existing approaches still lack formal quantification of the uncertainties that pervade the underlying data and any corresponding analysis of how the models reach their decisions.

The core of the innovation lies in a strategy the authors call geological-knowledge-constrained data augmentation, or GDA, applied as an inference test-time augmentation. Rather than feeding a single, potentially flawed dataset into a trained model, the GDA generates geologically plausible variations of the evidence layers, with each variation constrained by geological and geophysical understanding. The perturbations are therefore not random noise but realistic alternatives that respect what is already known about the subsurface environment, allowing the model to be interrogated under a range of credible data scenarios.

For each test sample, multiple augmented realizations are passed through the trained model, and the ensemble of forward predictions is aggregated to estimate prospectivity probability and to quantify data uncertainty. In practical terms, if many plausible versions of the input data all point to the same patch of rock being favorable for mineralization, confidence rises; if the predictions scatter widely across those versions, the framework flags that location as uncertain. This ensemble approach transforms a single deterministic output into a probabilistic one, giving exploration geologists a measure of reliability alongside the raw prediction itself.

Interpretability is handled through multi-forward-pass Shapley additive explanations, which evaluate the relative contributions of different evidence layers to the model’s outputs. By running the explanation analysis across the ensemble of augmented predictions, the framework reveals the decision-making basis of the 3D prospectivity model, showing which geological or geophysical inputs drove a given classification. This transparency matters because a model that merely labels a volume of rock as prospective, without indicating why, offers little guidance for refining exploration strategies or for judging whether the model has latched onto genuine geological signals or spurious correlations.

To separate promising targets from unreliable ones, the framework employs a risk–coverage curve, a tool that identifies subsurface areas characterized by both high prospectivity probabilities and low data uncertainty. This dual criterion is significant because a high probability alone can be misleading if the data supporting it are shaky. By explicitly coupling confidence in the prediction with confidence in the data, the method directs attention to targets where the model’s evidence is both strong and stable.

The backbone classifier for the trial is a 3D LeNet-5 model, an adaptation of a classic convolutional neural network architecture to volumetric data. The researchers trialed the framework in the Wulong gold district on the Liaodong Peninsula in China, a setting that provided the geological and geophysical evidence layers needed to exercise every component of the pipeline, from augmentation through uncertainty estimation to explanation.

The results, according to the authors, indicate that the proposed framework improves reliable and explainable 3D mineral prospectivity mapping through prospectivity probability estimation, uncertainty quantification, and evidence-layer contribution analysis. Each of these three outputs addresses a distinct failure mode of conventional mapping: point estimates that ignore uncertainty, black-box decisions that resist scrutiny, and target lists that mix robust predictions with fragile ones. By tackling all three within a single pipeline, the framework offers a template for how machine learning in the geosciences can be made both more trustworthy and more useful in practice.

The broader significance of the work extends beyond the Wulong district. Data uncertainty is inherent in 3D mineral prospectivity mapping because it arises from the uncertainties associated with 3D geophysical inversion and geological modeling, processes that inevitably involve assumptions, interpolation, and imperfect measurements. Embedding uncertainty awareness directly into the mapping workflow, as this framework does, gives decision-makers a clearer picture of where additional data acquisition would most improve target confidence. For the mining and exploration industry, the implications are tangible: drilling campaigns are expensive, and prioritizing targets that combine high prospectivity with low data uncertainty could reduce the risk of committing resources to poorly constrained volumes of rock.

Subject of Research: A data-uncertainty-aware and explainable framework for three-dimensional mineral prospectivity mapping

Article Title: A Data-Uncertainty-Aware and Explainable Framework for 3D Mineral Prospectivity Mapping

Article References: Wu, Z., Zhang, Z., Carranza, E. J. M., & Lu, Z. (2026). A Data-Uncertainty-Aware and Explainable Framework for 3D Mineral Prospectivity Mapping. Natural Resources Research. https://doi.org/10.1007/s11053-026-10793-x

Image Credits: AI Generated

DOI: 10.1007/s11053-026-10793-x

Keywords: mineral prospectivity mapping, 3D modeling, data uncertainty, explainable AI, Shapley additive explanations, test-time augmentation, geophysical inversion, Wulong gold district, machine learning, mineral exploration, risk–coverage curve, Liaodong Peninsula