Abstract: The fundamental challenge in deep mineral exploration lies in the extreme scarcity of known mineralization information in the target area, which severely restricts the generalization ability of traditional data-driven models. This study aims to develop a three-dimensional mineralization intelligent prediction framework based on internal knowledge transfer to overcome the prediction uncertainty caused by the scarcity of samples in deep areas (depth > 500 m). A unified three-dimensional geological model was constructed, and a three-dimensional graph structure integrating geological semantics was innovatively introduced to represent spatial topological relationships. On this basis, a GAT-JDA model capable of deep transfer learning was proposed. This model uses a graph attention network as the backbone architecture to capture the spatial association patterns of mineralization in the shallow, data-rich areas. Meanwhile, it integrates domain adversarial training and multi-kernel maximum mean discrepancy minimization strategies to form a joint domain adaptation mechanism, aiming to extract domain-invariant features shared by the shallow and deep areas, thereby achieving robust knowledge transfer from the shallow (source domain) to the deep (target domain). The model was fine-tuned using limited deep samples to generate a three-dimensional mineralization probability field in the deep area. The results show that compared with the baseline model using only a small number of deep samples, the F1 value of the GAT-JDA model on the independent test set significantly increased from 0.610 to 0.819, and the area under the curve reached 0.892. Feature distribution visualization analysis confirmed that this method can effectively align the feature spaces of the shallow and deep areas. Based on geological understanding and prediction results, three promising target areas for mineralization have been delineated, which can serve as the key areas for further exploration in the mining area.