An Interpretable Graph Attention Network for Mineral Prospectivity Mapping

Various data-driven mineral prospectivity mapping (MPM) methods have been successfully adopted to delineate prospecting regions for a specific type of mineral deposit. These methods are mainly developed based on pixel-wise or image (pixel-patch) data, which do not adequately consider the spatial patterns linked to mineralization or the spatial characteristics of mineral deposits to some extent. Graphs that are composed of nodes and edges have a strong ability to capture complex and nonlinear spatial coupling relationships. In addition, data-driven MPM methods typically ignore domain knowledge and expert experience, resulting in poor generalization ability and interpretability, and a lack of consistency in physical laws. In this study, an interpretable graph attention network (GAN) was proposed to map the mineral potential for Fe polymetallic mineralization in southwestern Fujian Province of China. Prior domain knowledge was added to the construction of an interpretable GAN. For the input data, a topology graph construction criterion based on prior geological knowledge was proposed to create graphs according to the degree of correlation between nodes. In the training process, a soft geological constraint was achieved by adding a penalty term to the loss function of the GAN based on a nonlinear relationship between the prospectivity density and distance to controlling features. Furthermore, a hidden layer was added to the network structure of the GAN to implement the hard geological constraint based on the ore-forming controlling features. A comparative study of an interpretable GAN and a conventional GAN demonstrated that the former can improve the probability in areas with high mineralization potential, and increase the interpretability of the obtained results.