Comparative analysis of multiple machine learning models: identifying impact factors in biochar heavy metal...

The contribution analysis of influencing factors governing biochar-mediated heavy metal adsorption in aqueous systems holds significant implications for cost-effective water remediation. Current studies predominantly rely on single-model approaches to identify critical variables, which may introduce bias due to inherent model assumptions, thereby impeding systematic elucidation of impact mechanisms and variable interactions. To address this gap, we integrated twelve machine learning models with SHAP (Shapley Additive exPlanations) interpretation to holistically investigate determinants and key variables. Modeling results demonstrate superior predictive accuracy and interpretability of tree-based architectures. SHAP importance analysis reveals initial metal concentration (C₀) exhibits the highest contribution (23.1-67.9%), followed by biochar factors (20.1-57.4%), Biosorption conditions (8.4-19.8%), with intrinsic metal properties showing minimal impact (< 5%). Core determinants are identified as C₀, cation exchange capacity (CEC), pH of solution (pH_sol), and pH of biochar surface ( ${\text{pH}}_{{\text{H}}_2\text{O}}$ ). SHAP dependence analysis further indicates: (1) elevated C₀ and CEC values substantially enhance adsorption capacity; (2) pH_sol exerts significant control under extreme acidity/alkalinity; (3) ${\text{pH}}_{{\text{H}}_2\text{O}}$ demonstrates inhibitory effects within pH 0-10 but markedly enhances performance at pH > 10. This multi-model ensemble framework overcomes single-model limitations, systematically deciphering the adsorption mechanism while quantifying variable contributions. The findings establish critical references for biochar design optimization and advance both theoretical understanding and practical implementation of heavy metal removal technologies.