Predictive modeling of heavy metal lead and cadmium adsorption on biochar based on machine learning

Biochar has great potential as an adsorbent for heavy metal ions, and predicting its adsorption performance using machine learning algorithms is a promising research area. This study employs two machine learning methods, lightweight gradient boosting machine (LightGBM) and deep neural network (DNN), to establish predictive models for the removal rates of Pb²⁺ and Cd²⁺ by biochar. The dataset for Pb²⁺ and Cd²⁺ contains 419 and 240 samples, respectively. Shapley Additive Explanations (SHAP) values are also used to analyze the role of functional groups in the adsorption of heavy metal ions. By comparing two input feature combinations (with/without elemental analysis), it was found that adding elemental analysis can improve the prediction accuracy of Pb²⁺ removal rate. Specifically, the R² of the LightGBM model increased from 0.920 to 0.923, and the MAE and RMSE were reduced by 0.761 and 0.641, respectively. However, the inclusion of elemental analysis showed little change in the prediction accuracy of Cd²⁺ removal rate. This study provides valuable insights for predicting biochar's adsorption of other heavy metal ions and further explains that different functional groups influence the adsorption performance of various heavy metal ions differently.