<p>Artificial intelligence (AI) for mineral prospectivity mapping (MPM) is a promising frontier in mineral exploration. However, due to the complexity of mineralization processes, the rarity of mineralization events, the diversity of mineralization features, and the black-box nature of AI, intelligent MPM faces several challenges, including inadequate representation of geological prospecting data and their spatial coupling relationships, insufficient training samples, poor model robustness, limited generalization capability, and lack of interpretability. This study systematically analyzes the underlying causes of these challenges in MPM and reviews previously proposed solutions. Accordingly, two novel AI models for MPM are proposed to address the above-mentioned issues: (1) a geologically constrained self-supervised Graph-Transformer model has the ability to mitigate the influence of limited labeled data by leveraging self-supervised learning. This model utilizes the graph structure to capture the spatial coupling between geological entities, and enhances the ability to model long-range spatial dependencies through the Transformer architecture; and (2) a geologically constrained graph reinforcement learning (RL) model can use the graph structure to represent geological features and comprehensively mine data through RL mechanisms. Additionally, geological knowledge is embedded into the reward mechanism of RL to incorporate mineralization knowledge into state discrimination, thereby enhancing its generalization ability and interpretability.</p>