Identification of Mineralizing Lithologies Driven by Interpretable Machine Learning on Multisource Geochemical Data: A Case Study from the Qin–Hang Metallogenic Belt

The Qin–Hang metallogenic belt (QHMB) constitutes a key Cu–Au polymetallic province in South China, yet deep exploration is hindered by weak geochemical anomalies and intense overprinting. This study established an interpretable dual-channel machine learning framework integrating whole-rock and zircon geochemical datasets to predict mineralization potential. Data from 71 representative mineralized sites comprising 46 whole-rock and 32 zircon geochemical variables were analyzed using nine supervised algorithms optimized via Optuna and Hyperopt with fivefold cross-validation and evaluating through accuracy, precision, recall, F1-score, AUC, and Kappa. LightGBM and CatBoost achieved the best performance, with accuracies of 0.95 for whole-rock and 0.91 for zircon datasets. SHAP interpretation identified Fe 2 O 3 T , Eu/Eu*, CaO, Zr/Hf, and Y as dominant whole-rock indicators and Eu, Er/Lu, Th, LREE-I, and Ce/Nd as principal zircon variables. Interaction effects (Eu/Eu*–Nd, Er/Lu–Tb/Tm, Er/Lu–Dy/Yb) indicate coupled controls of magma differentiation, redox state, and volatile evolution. Whole-rock geochemistry reflects shallow hydrothermal processes but is susceptible to alteration, whereas zircon retains deep-source signatures and resists post-magmatic overprinting, enabling early detection of concealed mineralization. These geochemical fingerprints further record the Mesozoic tectonomagmatic transition of the QHMB, during which crustal thickening followed by lithospheric extension generated oxidized and hydrous magmas favorable for porphyry–skarn Cu–Au mineralization. The dual-channel framework enhances predictive robustness and accuracy, providing a cost-effective, interpretable tool for deep mineral exploration in the QHMB and analogous metallogenic systems.