A Framework for Data-Driven Mineral Prospectivity Mapping with Interpretable Machine Learning and Modulated Predictive Modeling

Although mineral prospectivity modeling (MPM) has undergone decades of development, it has not yet been widely adopted in the global mineral exploration industry. Exploration geoscientists encounter challenges in understanding the internal working of many mineral prospectivity models due to their black box nature. Besides, their predictive results usually delineate undesirably large high-prospectivity areas, which are biased toward existing deposits, making MPM impractical. However, there are only a few data-driven methods for MPM that address both the interpretability of black box models and the issue of bias in high prospective areas, which may result from the intrinsic stochastic uncertainty of training samples, particularly toward well-known deposits. In this study, we construct and demonstrate a framework to improve the performance and reliability of data-driven MPM in the Qulong-Jiama mineral district of Tibet. Firstly, the mineral systems concept was applied to select appropriate targeting criteria and to derive corresponding evidential features. Secondly, model-agnostic methods, such as permutation feature importance, partial dependence plot, individual conditional expectation plot, and Shapely values, were applied to interpret the machine learning models. Finally, modulated prediction models and the spatial pattern of linked uncertainties were generated by an ensemble method that combines bootstrapping and the Random Forest algorithm. The final exploration targets, which were demarcated by cells with high modulated values and low uncertainties obtained by 50 predictive models, account for just similar to 3% of the study area.