R-GeoXNet: A Direction-Aware Relational Graph Neural Network for Regional Mineral Prospectivity Mapping

Abstract Mineral prospectivity mapping in structurally complex terranes requires models capable of capturing anisotropic geological controls, including fault-guided fluid flow, intrusive contacts, and directional geochemical dispersion. This study proposes R-GeoXNet, a direction-aware relational graph neural network for regional polymetallic prospectivity mapping in the Lhasa-Woka area of the eastern Gangdese metallogenic belt, Tibet. The framework integrates 39 stream-sediment geochemical elements, three isometric log-ratio balances, and fracture-density information into 16 × 16 raster tiles, which are transformed into pixel-level graphs with nine spatial relations representing orthogonal, diagonal, and self-loop connectivity. Direction-aware relational message passing learns geochemical-structural patterns linked to mineralization. Using 234 labeled tiles for training and evaluation and 5,500 tiles for regional prediction, R-GeoXNet achieved 85.42% accuracy, 0.9306 ROC-AUC, 0.934 PR-AUC, and 0.8444 F1-score, outperforming GCN and GAT baselines.