Exploration
Abbas Bahroudi; Salman Farahani
Abstract
The increasing depletion of near-surface ore deposits and the growing complexity of subsurface geological environments have intensified the need for data-driven, three-dimensional frameworks in mineral exploration. This study introduces an integrated 3D ore prospectivity modeling approach that combines ...
Read More
The increasing depletion of near-surface ore deposits and the growing complexity of subsurface geological environments have intensified the need for data-driven, three-dimensional frameworks in mineral exploration. This study introduces an integrated 3D ore prospectivity modeling approach that combines a Deep Autoencoder (DAE) with Monte Carlo Dropout (MCD)-based uncertainty quantification to generate both high-resolution prospectivity predictions and robust estimates of model confidence. A multi-source geoscientific dataset—comprising geology, geochemistry, geophysics, and borehole information—from the Siahcheshmeh intrusion-related gold system in northwestern Iran was voxelized into a unified 3D grid. The multi-scale convolutional DAE architecture effectively learned latent spatial patterns associated with alteration zones, structural intersections, and geophysical anomalies, while 50 stochastic forward passes via MCD enabled the decomposition of aleatoric and epistemic uncertainties. The proposed DAE–UQ model achieved an accuracy of 96.8% and an ROC-AUC of 0.96, outperforming conventional autoencoders, CNNs, and Random Forest models by 4–5%. High-prospectivity regions (>0.72) accounted for only 24% of the model volume yet captured 68% of mineralized borehole intercepts. Uncertainty analysis revealed elevated uncertainty at the margins of data-sparse zones, and excluding high-uncertainty voxels increased prediction accuracy to 98.6%. The spatial correspondence between high-prospectivity voxels, Au–Cu anomalies, silicification halos, and transpressive fault systems validates the geological reliability of the model outputs. Overall, the DAE–UQ framework offers a scalable, uncertainty-aware solution for 3D mineral prospectivity analysis in structurally complex metallogenic terrains. Its strong generalizability and robustness highlight its potential for application to other deposit types and emerging multi-source geoscience datasets.
