DATA-DRIVEN MODELING OF ORGANIC POLLUTANT ADSORPTION ONTO ADVANCED ADSORBENT MATERIALS UNDER VARIABLE PHYSICOCHEMICAL CONDITIONS

Abstract

Wastewater treatment is a persistent problem with organic pollutant contamination due to the significant role that the interacting material, pollutant, physicochemical, and operational factors play in the removal efficiency. This study has created a data-driven prediction model for advanced adsorbent material adsorption capacity under changing treatment conditions. An initial 6,172 observations were analyzed, with 101 duplicate observations removed, leaving 6,071 unique observations to be analyzed. Eight variable factors: specific surface area, pore volume, molecular characteristic, pH, temperature, initial concentration, flow rate, and column height were studied with respect to the adsorption capacity (Q). The regression approaches compared were: Multiple Linear Regression, Random Forest, XGBoost, Support Vector Regression, and K-Nearest Neighbors, with an 80:20 train-test split and 5-fold cross-validation. XGBoost demonstrated the best overall performance with R² = 0.975 and RMSE = 12.529 mg/g and the lowest MAE of 3.842 mg/g was obtained with Random Forest. Initial concentration is the most important feature in XGBoost, making up 65.404% of its importance, followed by column height and specific surface area. However, SHAP analysis revealed that the adsorption capacity increased with increasing concentration initially, and other variables had nonlinear and context-dependent effects. The results show the capability of interpretable machine learning models to predict the adsorption process accurately, optimize it, and minimize the number of experiments required for wastewater treatment. The outcomes also highlight the importance of considering both structural and operating parameters together.