Machine-learning-based prediction and optimization of emerging contaminants' adsorption capacity on biochar materials

Biochar materials have recently received considerable recognition as eco-friendly and cost-effective adsorbents capable of effectively removing hazardous emerging contaminants (e.g., pharmaceuticals, herbicides, and fungicides) to aquatic organisms and human health accumulated in aquatic ecosystems. In this study, ten tree-based machine learning (ML) models, including bagging, CatBoost, ExtraTrees, HistGradientBoosting, XGBoost, GradientBoosting, DecisionTree, Random Forest, Light gradient Boosting, and KNearest Neighbors, have been built to accurately predict the adsorption capacity of biochar materials toward ECs in aqueous solutions. A very large data set with 3,757 data points was generated using 24 input variables (i.e., pyrolysis conditions for biochar production (3 features), biochar characteristics (3 features), biochar compositions (6 features), and adsorption experimental conditions (12 features)) obtained from the batch adsorption experiments to remove 12 kinds of ECs using 18 different biochar materials. The rigorous evaluation and comparison of the ML model performances shows that CatBoost model had the highest test coefficient of determination (0.9433) and lowest mean absolute error (4.95 mg/g), outperformed clearly all other models. The feature importance analyzed by the shapley additive explanations (SHAP) indicated that the adsorption experimental conditions provided the highest impact on the model prediction for adsorption capacity (41 %) followed by the adsorbent composition (35 %), adsorbent characterization (20 %), and synthesis conditions (3)%). The optimized experimental conditions predicted by the modeling were a N/C ratio of 0.017, BET surface area of 1040 m2/g, content of C(%) contents of 82.1 %, pore volume of 0.46 cm3/g, initial ECs concentration of 100 mg/L, type of pollutant (CAR), adsorption type (Single) and adsorption contact time (720 min).