Physics-Informed Neural Network Boosts Mineral Prospectivity Modeling

A new physics-informed neural network, named Korzhinskii-Net, has been introduced to enhance mineral prospectivity modeling. This model integrates fundamental subsurface physical processes such as heat advection, fluid flow, and lithology-dependent precipitation, which are crucial for localizing ore deposits. Unlike conventional data-driven classifiers that rely on shallow surface proxies, Korzhinskii-Net leverages a deeper understanding of geological physics. The network is a 2-D radial architecture that couples Darcy flow, advective-diffusive heat transport, and a reaction rate into a single differentiable forward model. It is weakly supervised using readily available surface and remote-sensing data. Extensive evaluation across five diverse ore provinces and four commodity types demonstrated its superior performance, achieving a mean PR-AUC of 0.885 compared to 0.281 for the best classical baseline. This significant improvement suggests that incorporating physics-informed differentiable simulators can uncover localization patterns that purely feature-based learning approaches often miss, even when constrained by global open-data proxies. The full pipeline and evaluation harness have been released as open source.