Enhanced machine learning prediction of biochar adsorption for dyes: Parameter optimization and experimental validation

Abstract Biochar, as an eco-friendly, carbon-rich,and economical adsorbent, proven effective in removing toxic dyes from aquatic environments. This study evaluated the efficacy of machine learning (ML) models in predicting the adsorption capacity of biochar for dye removal. Nine models, namely CatBoost, XGBoost, Gradient Boosted Decision Trees, Random Forest, Histogram-Based Gradient Boosting, Kernel Extreme Learning Machine, Kriging, Light Gradient Boosting Machine, and AdaBoost, were deployed to ascertain their predictive accuracies. The CatBoost model was highlighted for its exceptional performance, achieving the highest R 2 (0.9880) and the lowest RMSE (0.0839). The stability of the model was affirmed through residual analysis and random partitioning dataset. A detailed feature importance analysis revealed that experimental conditions predominantly affect adsorption, accounting for 50.8%, followed by biochar characteristics (34.1%) and dye types (15.1%). The most significant feature impacting dye adsorption was identified as the C 0 through SHapley Additive exPlanations. Partial dependence plots were used further to illustrate the influence of features on the predictive model. Additionally, experimental validation of the ML approach yielded R 2 of 0.9037, reinforcing the applicability of the model. This study adds to supportive evidence of the use of ML for the prediction of adsorption capacity and encourages the development of user-friendly software, using PySimpleGUI, opening new paths to advanced data-driven methods in environmental engineering. Graphical Abstract