Document Type : Original Research Paper
Authors
1 School of Mining, Petroleum and Geophysics, Shahrood University of Technology, Shahrood, Iran
2 Shahrood University of Technology
Abstract
Mineral prospectivity modeling in structurally complex and vertically heterogeneous geological systems requires analytical frameworks capable of capturing nonlinear feature interactions and depth-dependent variability. This study evaluates the predictive performance of a deep self-attention neural network within a fully 3D mineral prospectivity modeling framework applied to the Chah-Mousa copper deposit, Iran. The modeling domain was discretized into twenty-one independent elevation levels to assess depth-consistent predictive behavior. Model performance was evaluated using ROC–AUC analysis, confusion-matrix-derived metrics, and success-rate curve assessment. The deep self-attention model achieved a mean ROC–AUC of approximately 0.83, indicating strong discriminative capability between mineralized and non-mineralized domains. Averaged across elevation slices, classification performance remained stable (Accuracy ≈ 0.83, Precision ≈ 0.69, Recall ≈ 0.75, F1-score ≈ 0.72), demonstrating vertical generalization and resistance to shallow overfitting. Success-rate analysis revealed that more than 50% of known mineralized occurrences are concentrated within the top 10% of predicted prospectivity areas, confirming strong ranking efficiency for exploration prioritization. The probabilistic outputs exhibit spatial coherence aligned with structural corridors and alteration zones, indicating that the attention mechanism effectively captures nonlinear geological relationships. The results demonstrate that deep self-attention architectures provide statistically robust, depth-consistent, and operationally meaningful predictions for 3D mineral exploration targeting in structurally controlled copper systems.
Keywords
Main Subjects
)