Machine learning-assisted modeling and analysis of PFAS removal from contaminated water via membrane-based treatment

Perfluoroalkyl and polyfluoroalkyl substances (PFAS) have garnered worldwide concern owing to their pervasive environmental persistence and deleterious effects on aquatic resources. Membrane separation technologies, particularly nanofiltration (NF) membranes and reverse osmosis (RO) membranes, have potential in PFAS removal, but the multi-factor coupling effects of these technologies still lack systematic research. This study developed sparrow search algorithm (SSA) and AdaBoost algorithms optimized kernel extreme learning machine (KELM) model to analyze the performance of 24 types of RO and NF membranes for removing 17 different PFAS from contaminated water. Based on the PFAS removal mechanism of RO and NF membranes and the operating conditions, twelve characteristic variables were selected as model inputs. Compared with the partial least squares, ridge regression, random forest, gradient-boosting, XGBoost, support vector machine, and BP neural network, the baseline KELM model delivered competitive accuracy (R²=0.56, RMSE=10.31). After SSA-AdaBoost optimization, predictive performance rose significantly (R²=0.85, RMSE=5.04). SHapley Additive exPlanations algorithm and partial dependence plots were applied to investigated the feature importance and the coupling effect of characteristic variables on the membrane removal of different PFAS. The results identified PFAS molecular weight, pH, and membrane pore size as the three dominant factors governing rejection. Optimum performance could be attained when the membrane exhibited a pore size < 0.4 nm, zeta potential ≈ -28 mV, surface roughness > 82 nm, and contact angle > 28°. By integrating these quantitative relationships, the study delivers a data-driven theoretical framework for tuning NF/RO membranes to optimize PFAS separation efficiency.