Explainable Machine Learning Combined with Feature Selection and Bayesian Optimization for Predicting Rare Earth Elements and Yttrium in Coal: A Case Study of China’s Five Major Coal-Bearing Regions
来源：ACS Publications
Determining the concentrations of rare earth elements and yttrium (REY) in coal is essential for evaluating coal-hosted critical metal resources and understanding their geochemical controls. However, conventional laboratory analyses are costly, time-consuming, and spatially sparse, limiting regional-scale assessment. In this study, we compiled 527 coal samples from 30 coalfields across China’s five major coal-bearing regions and developed an explainable machine learning framework for REY prediction. Unlike previous site-specific models, this framework integrates feature selection, Bayesian optimization, and SHAP interpretation for REY prediction at a broad regional scale. After data preprocessing, multicollinearity diagnosis, and correlation-based feature selection, 16 predictors were retained for model construction. Four ensemble regression models, including Random Forest (RF), Extra Trees (ET), CatBoost (CAT), and Gradient Boosting Decision Tree (GBDT), were optimized using Bayesian hyperparameter search and 5-fold cross-validation. The ET model achieved the best test-set performance among the four ensemble models, with RMSE = 39.962, MAE = 29.669, and R2 = 0.716, although predictions for highly enriched samples remained uncertain because of the long-tailed REY distribution. SHAP analysis identified Ga as the most influential predictor, followed by geological age and P2O5, suggesting that clay-related terrigenous input, stratigraphic context, and minor phosphate-associated phases may jointly influence model-predicted REY enrichment. These results demonstrate that the proposed framework can serve as an effective prescreening tool to prioritize coal samples for detailed REY analysis, while independent validation is recommended for high-value target areas.