<p>Deep learning has, in recent years, seen widespread application in the field of 3D mineral prospectivity mapping, leading to significant interest in data-driven 3D methods. However, this field still faces two critical challenges. On the one hand, most existing methods rely on convolutional neural network (CNN), but the locality of convolution operations limits their ability to capture long-range dependencies within 3D geological data. This shortcoming severely impacts model performance, particularly when handling anisotropic 3D data. On the other hand, the “black-box” nature of deep learning models results in a lack of interpretability regarding their decision-making processes. To address these issues, this paper proposes a hybrid architecture integrating 3D CNN and Transformer (Trans) and introduces a spatial continuity loss to construct a deep learning model with 3D gradient feature constraints (Geo CNN-Trans). This model combines the strengths of local feature extraction and global contextual modeling, while utilizing visualized gradient heatmaps to verify its effectiveness in capturing global relationships within the data. Using the Zaozigou mining area in Gansu Province, China, as a case study, the proposed Geo CNN-Trans model is compared with 3D CNN and CNN-Trans. The results demonstrate that Geo CNN-Trans excels at handling the anisotropy inherent in 3D data and can accurately identify global relationships within the 3D geological space. This approach provides a more accurate and reliable framework for 3D mineral prospectivity prediction.</p>