A Hybrid 3D CNN-GAT Framework with Entropy-Guided Adaptive Fusion for 3D Mineral Prospectivity Modeling

Reliable three-dimensional (3D) mineral prospectivity modeling for concealed ore targeting remains challenging because voxel-based models emphasize local volumetric patterns, whereas nonlocal spatial relationships among sampled locations and prediction reliability are not jointly represented. To address this problem, this study developed a hybrid 3D CNN-graph attention network (GAT) framework for the Anqing skarn Cu district, eastern China. A ResNet3D-S&E branch extracts local structural patterns from 7 × 7 × 7 voxel patches, while a GATv2 branch represents coordinate-based neighborhood relationships among sampled voxels. Predictive entropy supports entropy-guided adaptive fusion and reliability-aware interpretation rather than full Bayesian uncertainty quantification. The full model achieved an area under the prediction rate curve of 0.942 and an area under the receiver operating characteristic curve of 0.964 on the validation dataset. The probability volume was converted into coherent exploration targets through smoothing, connected-component analysis, vertical-continuity filtering, geological-association screening, and spatial merging. Five distal prediction targets and one known-mineralization-supported reference target were delineated, linking 3D prospectivity prediction, predictive entropy, and geology-guided target interpretation.