Quantifying uncertainty of mineral prediction using a novel Bayesian deep learning framework

Mineral resource exploration increasingly demands not only accurate prospectivity maps but also reliable measures of confidence to guide high-stakes decisions. In this study, a novel Bayesian deep learning (BDL) framework was introduced, which embeds probabilistic inference within a deep neural network to jointly predict mineralization potential and quantify uncertainty. Two posterior approximation strategies, Metropolis–Hastings (MH) sampling and variational inference (VI), are implemented to estimate model weights as distributions rather than as fixed values, enabling decomposition of predictive uncertainty into aleatoric and epistemic components. When applied to eleven ore-controlling features in the Nanling tungsten polymetallic region (China), both MH-based and VI-based BDL models demonstrate strong classification performance while revealing contrasting spatial patterns and uncertainty patterns. Correlation studies across probability bands confirm that MH sampling captures a broader spread of uncertainty at the cost of greater computational demand, while VI delivers greater efficiency but risks underestimating uncertainty. The results highlight trade-offs between accuracy, interpretability, and computational load, demonstrating that MH-based BDL offers more robust uncertainty assessments, whereas VI-based BDL places greater emphasis on efficiency. By providing spatially explicit probability and uncertainty maps, this framework advances risk-aware mineral exploration, enabling practitioners to target areas of high potential with low uncertainty and to identify regions warranting additional data acquisition. • Developing a novel BDL framework for uncertainty-aware MPM • Separating uncertainty into aleatoric and epistemic components • Quantifying predictive uncertainty in mineral exploration • Comparative analysis of Metropolis-Hastings and variational inference