A dual-track machine learning framework for andesite tectonic environment identification: ensemble learning and few-shot learning

Accurate discrimination of andesite tectonic settings is critical for unraveling Earth’s geodynamic processes. However, existing studies face three key challenges: (1) simplified traditional methods, which rely on single-element ratios and fail to capture the complex petrogenetic processes of andesites; (2) poor performance on small samples, as rare tectonic types (RV) are often misclassified owing to data scarcity; and (3) limited geological interpretability, with most models lacking clear links between geochemical features and magmatic mechanisms. To address these issues, we propose a “dual-track” framework integrating machine learning and few-shot learning using 26,463 andesite samples from the GEOROC database. For large-sample scenarios, optimized ensemble models (Random Forest, XGBoost, LightGBM) achieve high precision, with an Area Under the Receiver Operating Characteristic Curve (AUC, a metric reflecting overall classification performance) ≥ 0.99. LightGBM emerges as the dominant model, with a recall rate of 97% for small-sample RV. For rare tectonic types, a meta-learning (TabPFN pre-training) and knowledge distillation (transfer to CatBoost) framework boosts the recall rates of RV and OI to 99% while optimizing the inference speed to 0.01 seconds per sample. SHAP analysis identifies key discriminant elements (e.g., TiO2 and FeOt for CM; Nb and Lu for OI) and their synergistic effects, verifying classical magmatic theories (e.g., Fe-Ti oxide differentiation in subduction zones). This framework provides a reproducible standard for intermediate igneous rock classification, aiding paleotectonic reconstruction and mineral exploration in the future.