Korzhinskii-Net Delivers 0.885 Mean PR-AUC Across Five Ore Provinces Using Physics-Informed Neural Networks

The arXiv preprint from Boris Kriuk presents a weakly supervised physics-informed network that embeds infiltration metasomatism equations directly into the loss function. Training relies solely on global open-data surface proxies and remote-sensing layers yet recovers subsurface localization patterns that pure data-driven classifiers miss. Five-fold cross-validation with hard ring-shaped negatives yields consistent gains across nickel, copper, gold, and diamond systems.

Classical machine-learning approaches treat prospectivity as a static classification task on surface features. Korzhinskii-Net instead solves a differentiable forward model of fluid flow and temperature-dependent precipitation, allowing gradients to propagate through the governing physics. This produces fractional ranks of 0.019 versus 0.413 for the strongest baseline, indicating that the network ranks known deposits far higher even when surface expressions are subtle.

The work connects directly to earlier PINN successes in reservoir simulation and seismic inversion but applies the framework to mineral systems for the first time at province scale. Releasing the full pipeline enables immediate replication and extension to 3-D models once higher-resolution geophysical constraints become available. Operational adoption will require integration with company proprietary drill-hole and downhole geophysical datasets to move beyond the current proxy-only supervision.

Next steps include coupling the network to stochastic inversion frameworks and testing on greenfield terranes where no training labels exist. If successful, such models could shorten exploration cycles by prioritizing drill targets that satisfy both data fit and physical plausibility.