Machine Learning-Guided Prediction of the Removal Efficiency of Pharmaceuticals and Personal Care Products by Photocatalytic Membranes

Pharmaceuticals and personal care products (PPCPs) are contaminants of emerging concern that threaten both ecosystems and human health. Photocatalytic membranes (PMs), which combine photocatalysis with membrane filtration, have shown promise for the removal of PPCPs. However, their performance depends on multiple factors, including catalyst and membrane characteristics, operational conditions, and PPCP properties. In this study, machine learning (ML) models were developed to predict PPCP removal efficiency by PMs using a data set of 339 data points covering 74 PPCPs, 6 photocatalyst categories, and 9 membrane materials, compiled from 101 peer-reviewed studies. Among the tested models, the gradient boosting regression tree (GBRT) achieved the best performance, with an R2 of 0.645, a mean absolute error of 0.110, and a root-mean-squared error (RMSE) of 0.151. Feature importance analysis using SHapley additive exPlanations (SHAP) revealed that operational conditions (e.g., light time, hybrid methods, and initial concentration) and photocatalyst category were dominant factors. Substructure analysis further identified hydroxyl and heterocyclic groups as critical determinants of the PPCP degradation. These results demonstrate the value of ML in predicting the removal efficiency and providing new insights into the relationships among PPCP properties, photocatalyst characteristics, membrane materials, and operational conditions, offering guidance for the design and optimization of PM systems.