A portable source-informed machine learning framework for predicting heavy metals in legacy smelting sites with...

Legacy smelting sites often exhibit strong point-source control, highly right-skewed concentration distributions, and pronounced vertical stratification, which together challenge reliable spatial prediction of heavy metals (HMs) from limited borehole data. Here, we investigated a decommissioned Sb smelter in Southwest China and developed a pollution source information (PSI)-driven machine learning framework for Sb, Pb, and As. PSI was encoded from 29 representative sources as distance-based and distance-decay features, and model generalization was evaluated using a 5-fold GroupKFold, grouped by boreholes to mitigate spatial leakage. XGBoost was used as the primary algorithm and benchmarked against RF, SVM, and MLP. To improve predictive accuracy under heavy-tailed targets, an adaptive logarithmic target transformation (ALTT) and a hybrid sample-weighting scheme (HSWS) were introduced. Under leakage-resistant evaluation, the PSI-driven XGBoost models achieved out-of-fold R² values of 0.82 (Sb), 0.72 (Pb), and 0.71 (As), substantially outperforming models without PSI (R² = 0.28, 0.14, and -0.013, respectively). ALTT combined with HSWS further improved predictive performance. Incorporating source-distance and distance-decay features significantly enhanced model performance. Together with observations across multiple environmental media and SHAP interpretation, the results revealed shared control of HM distributions by source proximity, depth, and stratigraphy. However, metal-specific patterns remained evident: Sb and As were more source-driven, Pb was dominated by the soil layer, and As exhibited stronger multi-source superposition and directional gradients. This framework relies on routinely available site investigation data, delivers high predictive accuracy, and provides a practical reference for point-source-dominated legacy sites.