An explainable machine learning framework for predicting colloid-mediated release of potentially toxic elements from Pb/Zn smelter contaminated soils

To date, how the physicochemical properties of soil colloids contribute to the prediction and explanation of PTEs release remain unclear. Therefore, the machine learning (ML) model framework was developed and optimized to accurately predict PTEs concentrations in soil colloid solutions. The results indicated that colloid concentrations were within the order of 10¹⁰ particles/mL, with the zeta potential ranging from -49.5 to -5.60 mV, and the median particle sizes (D50) ranging between 190 and 200 nm. The mean PTEs concentrations followed the decreasing order of Zn (68.23 mg/L) > Cd (2.98 mg/L) > Pb (0.78 mg/L) > As (0.12 mg/L). A two-stage model coupled with random forest (RF) classification and gradient boost regression tree (GBRT) regression could greatly improve As and Pb prediction accuracy (R²>0.85). The extreme gradient boosting (XGBoost) model combining with logarithmic transformation achieved the higher predictive performance for Cd (R²=0.97). XGBoost further incorporating Winsorize processing and logarithmic transformation demonstrated the better predictive performance for Zn (R²=0.97). Relative importance analysis further indicated that soil types and colloid properties were the most influential factors for PTEs release. Collectively, these findings could provide the technical support for the risk prediction and management of PTEs in contaminated sites.