A deep-learning-based mineral prospectivity modeling framework and workflow in prediction of porphyry-epithermal mineralization in the Duolong ore District, Tibet

Machine learning (ML) is emerging as a highly effective technique for mineral exploration. However, mineral exploration poses several unique challenges to ML application, such as uncertain geological information in remote regions and imbalanced labeled training data. In this study, we developed a deep-learning framework - a self-attention back-propagation neural network (SA-BPNN) - which is used to automatically explore re-lationships among diverse features and improve the capability of information extraction. Moreover, we proposed a mineral prospectivity modeling workflow involving "quantitative data + ML + expert experience" for porphyry-epithermal deposits. Using quantitative data obtained from hyperspectral remote sensing, geochem-istry, and geophysics, we predicted ore-prospecting targets by applying the SVM, SA-BPNN, and U-Net models. Thereafter, we combined the model-based prediction with geological data to delineate the target areas. The model-based prediction by SVM, SA-BPNN, and U-Net occupy 1.73%, 1.40%, and 2.21% of the study area and contain 100%, 100%, and 80% of the known Cu-Au mineralization in the Duolong ore district in Tibet, respectively. The proposed SA-BPNN method, thus, achieved superior performance for mineral prospectivity modeling compared with alternative methods.