Mineral prospectivity mapping of porphyry copper deposits in the Duobaoshan district using random forest and SHAP interpretation

[Objective] Mineral resource prediction is often hindered by the complexity of metallogenic processes and the challenge of fusing multi-source geological data. To address these issues, the Duobaoshan copper deposit and its surrounding area in Heilongjiang Province were selected as a case study, where machine learning algorithms were applied for the prediction and evaluation of porphyry copper deposits. [Methods] By integrating multi-source geological data, a predictor system of eight factors was constructed, including buffers for faults, intrusions, and strata; geochemical anomalies of Cu, Mo, and Au; the first robust principal component score (RPC1); and residual gravity anomalies. To address the scarcity of known deposits, a spatial neighborhood augmentation strategy was adopted for sample expansion. On this basis, a Random Forest (RF) prediction model was developed, with Logistic Regression (LR) and Support Vector Machine (SVM) introduced as baseline models for performance comparison. Furthermore, the SHAP algorithm, utilizing the TreeExplainer and interaction plots, was employed to quantitatively interpret key metallogenic elements. [Results] Experimental results indicate that the grid-search optimized RF model achieved an AUC of 0.962 on the testing set, outperforming SVM (0.938) and LR (0.874), demonstrating superior generalization and robustness. Success-rate analysis showed that the top 10% high-probability area captured 88% of known deposits, indicating significant exploration efficiency. SHAP analysis revealed that stratigraphic buffer, RPC1, and Cu anomalies were the dominant predictors. Moreover, significant non-linear interaction enhancement effects were identified between strata and faults/Cu anomalies, quantitatively characterizing the synergistic metallogenic mechanism of "strata-structure-fluid". [Conclusion] This study constructed a random forest prediction model based on sample augmentation and multi-model comparison, effectively overcoming the difficulty of small-sample modeling. Based on probability thresholds determined by the success-rate curve, seven metallogenic prospective zones were delineated, including one Grade-A, four Grade-B, and two Grade-C zones. The prediction results are highly consistent with geological laws, providing scientific basis and technical support for the exploration of porphyry copper deposits in the Duobaoshan periphery and similar covered areas.