GeoGraphFormer: A Direction-Aware Graph Transformer for Regional Polymetallic Mineral Prospectivity Mapping

Abstract Regional mineral prospectivity mapping increasingly utilizes graph-based deep learning to integrate geochemical and structural information. Existing graph models, however, do not explicitly represent orientation-dependent geological relationships, which limits their capacity to preserve structurally controlled mineralization patterns. To address this limitation, GeoGraphFormer, a direction-aware Graph Transformer for regional polymetallic mineral prospectivity mapping, is introduced. This framework incorporates GeoRelAttn, which learns directional attention biases across nine spatial relations, and GeoBandWeighting, which adaptively weights 43 geochemical-structural input layers. When evaluated on the Lhasa-Woka metallogenic belt, GeoGraphFormer achieved 91.67% accuracy, 0.9444 ROC-AUC, 0.9478 PR-AUC, and a 0.9167 F1-score, outperforming Graph Convolutional Networks and Graph Attention Networks. These results indicate that explicitly encoding geological directionality enhances graph representations of structurally controlled mineralization and establishes a geologically informed framework for regional mineral prospectivity mapping.