DOI: https://doi.org/10.36347/sjet.2026.v14i10.001

Pages: 575-598

Mineral prospectivity mapping is difficult in regions where confirmed deposits are few and available geological data vary widely in type and quality. Standard machine learning and convolutional methods have improved prediction accuracy, but they ignore the directional structure of geological space, do not enforce geological logic during training, and rarely explain which factors drove a prediction. We present GeoEDL, a directional relational graph network built on geological constraints, combined with contrastive pretraining, for porphyry-gold prospectivity mapping under scarce labels. The framework provides interpretability at four levels: data, model architecture, geology, and results. Tiles are first classified as favorable, ambiguous, or barren using Mahalanobis distance and fracture density, without relying on mineralization labels, and these classes generate contrastive pairs for pretraining. A nine-edge-type graph encoder is then pretrained with geological contrastive loss and fine-tuned using evidential deep learning under structural and geochemical consistency penalties. On 9,324 tiles with 94 positive labels from Tuwu-Yandong, the model reached ROC-AUC 0.9418 and sensitivity 1.0000, with geochemical anomaly identified as the leading driver (51.6%) and epistemic uncertainty mapping flagging priority drilling targets.