Abstarct: Mineral potential modeling is regarded as one of the key yet inherently complex stages in the mineral exploration process, aiming to identify and prioritize prospective zones with minimum cost, time, and risk; however, due to the multi-source nature of exploration datasets, the presence of nonlinear relationships among geological evidences, and inherent uncertainties, this process has consistently faced major challenges, such that classical statistical and knowledge-driven methods—particularly in complex data-driven environments and three-dimensional analyses—exhibit limited capability in accurately representing these relationships. In this dissertation, a novel and integrated frxamework is proposed to enhance mineral potential modeling, baxsed on the combination of data-driven methods and uncertainty management; to this end, three data-driven algorithms, including the Bayesian method, Support Vector Machine, and a Deep Self-Attention model, were employed to extract complex spatial–structural patterns. Training data were prepared by converting mineralized polygons into dense point datasets and generating non-mineral points baxsed on a distance-baxsed logic to prevent data leakage and model bias, while the modeling was conducted in a three-dimensional manner through multi-level analysis across 21 independent elevation levels to precisely investigate vertical variations in mineral potential. Performance evaluation results indicated that among individual models, the Deep Self-Attention model achieved the highest discrimination power between mineralized and non-mineralized areas with a ROC–AUC value of approximately 0.83, whereas the Support Vector Machine and Bayesian models yielded AUC values of about 0.79 and 0.71, respectively. In the final step, integrating model outputs using Dempster–Shafer theory led to a significant improvement in all evaluation metrics, increasing the AUC of the ensemble model to approximately 0.89; moreover, the ensemble model successfully identified about 68% of known mineral occurrences within only the top 10% of the highest-potential area, compared to approximately 32% for the Bayesian model. The simultaneous increase in Recall (around 0.81) and Precision (around 0.76) indicates a reduced risk of missing potential deposits and improved control over low-efficiency drilling within the proposed frxamework. Overall, the results of this study demonstrate that combining data-driven models with the Dempster–Shafer frxamework, while reducing the inherent uncertainty of exploration data, significantly enhances the reliability of the final mineral potential map, and the proposed frxamework can be effectively applied as a practical tool for optimizing exploration decision-making, prioritizing drilling targets, and guiding exploration operations in the Chahmousa mine and other deposits with similar geological conditions.