Formulating gold prospectivity mapping as a constrained learning problem

We present a deep learning framework for rapid regional-scale gold prospectivity mapping using convolutional neural networks. Two predictive formulations are compared: (1) direct regression for gold concentrations, and (2) a novel distancebased approach that estimates proximity to high-grade mineralization, subject to specific constraints. Both models are trained and validated using data from the Yilgarn Craton in Western Australia, incorporating airborne geophysics and remote sensing datasets. We evaluate model performance using spatially distinct validation regions and highlight tradeoffs between local accuracy and spatial coherence. While the regression model directly fits gold assay values, the distancebased formulation offers smoother predictions and greater generalization in under-sampled areas. Our results demonstrate the flexibility of neural network based approaches for mineral targeting and underscore the importance of label formulation and domain-informed constraints in training models for sparse,spatially complex environments.