Mineral prospectivity prediction based on the dynamic relation model Atten-GCN: A case study of gold prospecting in the Yingfengjie area, Shaanxi province (northern China)

• In this study, the model’s predictive precision and interpretability are enhanced by integrating geological constraints, achieved through the computation of spatial correlations among geological features. • To more effectively capture the spatial characteristics of geological features, an Atten-GCN (Attention-based Graph Convolutional Network) was developed. • The synergistic integration of GCN and Attentionallows for the meticulous extraction of local geological feature details while simultaneously accommodating the overarching patterns of geological element variations. Mineral prospectivity mapping (MPM) based on deep learning commonly requires geological, geophysical, and geochemical data constraints. However, in areas with relatively complex spatial geological relationships, the lack of geological evidence layers could lead to poor interpretations in mineral prospectivity prediction. In this study, therefore, we propose to extract geological units and fault structures from geological maps as geological evidence layers, and geochemical survey data as geochemical constraints. With the Self-Attention mechanism, we train a mineral exploration prediction model based on GCN(Graph Convolutional Network) to perform MPM. To test this method, we conducted gold MPM in the Yingfengjie area (Shaanxi province, northern China). The results show that when the size of sampling window is set to 36 and the Atten relationship is set to 32, the prediction model achieves optimal performance with 96.59% accuracy. The predicted gold target area occupies 10.93% of the total district. Comparatively, our method yielded 0.81% and 2.13% higher accuracy than the prediction results by the GCN and CNN(Convolutional Neural Network) methods, respectively, demonstrating the superior prediction performance of this method. Compared to the gold prediction results using only geochemical data, the predicted area with geological evidence layers decreases by 7.47%, and the geological characteristics of the gold prediction area are highly similar to those of known deposits.