A generative neural network approach to uncertainty and risk-return analysis in mineral prospectivity modelling

Abstract

The industrial adoption of machine learning techniques for mineral prospectivity modelling (MPM) remains limited due to their inability to model uncertainties and a lack of systematic frameworks for evaluating risk and return in mineral predictions. A major challenge is that most existing methods fail to simultaneously capture both epistemic uncertainty, which arises from limitations in the predictive modelling process, and aleatoric uncertainty, which stems from the inherent randomness in geoscience data. To address this, we propose a conditional variational autoencoder (CVAE) approach incorporating decoder calibration and uncertainty estimation, which we apply to Canadian magmatic Ni (±Cu ±Co ±PGE) sulphide mineral systems. Aleatoric uncertainty is quantified from the CVAE’s posterior distribution, whereas epistemic uncertainty is assessed from 100 MPM realizations based on datasets generated by the CVAE. We also introduce a novel risk-return framework which integrates relative uncertainty measures with the non-parametric Getis–Ord $G^{\ast }$ statistics spatial clustering technique to categorize exploration targets into four distinct risk-return categories. Results from the spatial distribution and kernel density estimation analysis reveal that most known deposits are situated in low-uncertainty zones. Notably, high-return zones, which comprise approximately 4% of the total area, account for 94.7% of the known deposits. This research highlights the significance of incorporating uncertainty and risk-return analysis to improve decision-making in mineral prospecting.