A Geologically Constrained Variational Autoencoder for Mineral Prospectivity Mapping

Deep learning algorithms (DLAs) are becoming popular tools for mineral prospectivity mapping. However, purely data-driven DLAs frequently ignore expert and domain knowledge, imposing difficulty in interpretability from a geological perspective. The efficient integration of geological knowledge into DLAs remains challenging in geosciences. In this study, a geologically constrained variational autoencoder (VAE) was proposed to map prospectivity for gold mineralization in the Baguio District of the Philippines. A spatial nonlinear correlation between an ore-forming controlling feature and locations of mineral deposits was built as part of the loss function for constructing a geologically constrained VAE. A comparative study of a geologically constrained and a traditional VAE demonstrated that the former can enhance the probabilities in areas with high potential for locating mineralization and increase the interpretability of the obtained results.