Interpreting Machine Learning Predictions of Pb2+ Adsorption onto Biochars Produced by a Fluidized Bed System

Abstract

Employing machine learning to predict the Pb²⁺ adsorption capacity of biochars is an innovative pursuit in hazardous materials. This study compared artificial neural network (ANN), support vector regression (SVR) and random forest (RF) for Pb²⁺ adsorption capacity by biochar from a fluidized bed system. Besides developing correlations for comparison, the RF model (R² = 0.984, RMSE=0.054) outperformed both ANN (R² = 0.908, RMSE=0.316) and SVR (R² =0.667) in predicting higher adsorption capacity. Based on the superior performance, the Shapley Additive Explanations (SHAP) were employed on RF. SHAP global explanations indicated that adsorption conditions contributed 69.03% and biochar characteristics contributed 30.21%to adsorption capacity, highlighting Dosage (D) and Carbon (C) as the crucial factors. Regarding biochar characteristics, element compositions contributed 76.59%. The single samples demonstrated that the final predictions align with the experimental results. The synergistic effect of dependence plot explains the Pb²⁺ adsorption under varying parameter conditions, such as D<1g/L, C<45%, Pb_in>100mg/L, H<2.5, t>12h, T>25°C, pH>9, H/C>0.4, the SHAP value is positive, contributing to an increase in adsorption capacity. Furthermore, a graphical user interface (GUI) leveraging SHAP model parameters predicts adsorbent performance, providing novel insights into optimizing biochars production. The obtained findings narrow the search for optimal biochars adsorbents and might guide laboratory experiments and engineering application of Pb²⁺ removal using biochars.