<p>Mineral prospectivity mapping (MPM) is a pivotal tool for guiding mineral exploration by identifying high-potential target zones. While machine learning models have demonstrated remarkable predictive performance in MPM, most approaches often neglect critical aspects such as uncertainty quantification and model explainability, limiting their reliability and practical utility in high-risk exploration decision making. This study developed a novel uncertainty-aware deep learning (UADL) framework that integrates the Dirichlet distribution into deep learning networks to enhance uncertainty assessment and model explainability in mineral prospectivity analysis. The Dirichlet loss is introduced into the UADL to learn both accurate predictions and realistic uncertainty estimates. Besides, the UADL approach leverages Shapley additive explanations (SHAP) to provide interpretable, spatially explicit insights into the contributions of ore-controlling features. The UADL approach was validated through a Cu prospecting case study in the Zhongtiaoshan region (China), a region with complex ore-forming controls. The high areas under the curve for the training, testing, and full-area datasets confirmed that the UADL model performed consistently excellent and generalized effectively to unknown data. The model generated spatially coherent predictive map and uncertainty maps, which are critical for high-risk exploration decision making. The study indicates that the UADL approach provides robust capabilities and generality for mineral prospectivity analysis by merging Dirichlet distribution with deep learning.</p>