Graph attention network-based mineral prospectivity prediction: A case study of copper exploration in eastern Tien Shan, China
Shiyu Sheng, Yongzhi Wang, Jiangtao Tian, Xingyu Chen, Yan Ning, Yuhao Dong
Abstract
Traditional mineral prospectivity mapping (MPM) primarily relies on pixel-based analysis, which can exclude critical spatial patterns associated with mineralization. Through the attention mechanism, the Graph Attention Network (GAT) dynamically adjusts the weights between nodes, more accurately capturing the relationships between geological features. To enable intelligent copper prospectivity modeling, this study proposes integrating geological knowledge and graph neural networks. Using geological structure, metallogenic stratigraphy, and geochemical exploration data from the eastern Tien Shan orogenic belt, we constructed a unified data format and storage structure. We extract and fuse features related to copper mineralization, compile a training dataset, and apply the GAT algorithm to develop a comprehensive intelligent prediction model, Geoscientific Graph Attention Network (GeoGAT). Results indicate that the model effectively extracts features from multi-source geological data. It quantifies the suitability of prospective zones, achieving 83.3 % accuracy in identifying known copper deposits. Notably, it successfully detects numerous large known deposits. In comparative analyses with convolutional neural network (CNN) and multilayer perceptron (MLP) models, the GeoGAT model excels in assessing geological feature relevance and prediction accuracy. This study validates the effectiveness of the GAT algorithm in mineral prediction and highlights the application of graph attention networks as a frontier approach. Further exploration of their applications in mineral prediction is recommended.