Comparative Evaluation of Machine Learning Models for Micropollutant Removal Prediction in Membrane Bioreactor Systems

Membrane bioreactor (MBR) systems offer advanced capabilities for micropollutant attenuation, yet predicting their removal efficiency remains a significant challenge due to the physicochemical diversity of emerging contaminants. While machine learning (ML) models are increasingly applied to water treatment, comprehensive comparative evaluations utilizing extensive datasets and interpretable frameworks remain limited. This study evaluates six distinct ML algorithms including Linear Regression, Support Vector Regression, Artificial Neural Networks, Random Forest, Gradient Boosting, and XGBoost to predict the removal efficiency of 98 chemically diverse micropollutants. Utilizing a compiled dataset of 345 observations from MBR and osmotic MBR systems, model performance was rigorously assessed via extended metrics, including the Index of Agreement (IoA), Nash-Sutcliffe Efficiency (NSE), and Taylor diagrams. Tree-based ensemble models demonstrated superior predictive capacity, with XGBoost achieving the most robust performance (CV R² = 0.2853; RMSE = 9.6753%). To extract mechanistic insights from the optimal model, SHapley Additive exPlanations (SHAP) analysis was employed. The SHAP explainer identified that compound molecular weight and hydrophobicity (LogD) are the most dominant physical features dictating removal efficacy, heavily outweighing system parameters such as membrane type. Ultimately, the integration of XGBoost with SHAP provides a highly transparent, data-driven framework for understanding contaminant behavior, offering critical predictive insights for the optimization of MBR operating parameters in advanced wastewater treatment.