Coupling photocatalytic membrane engineering with machine learning for predictive fouling management in real pharmaceutical wastewater

Pharmaceutical effluents contain persistent organic pollutants that require advanced treatment solutions supporting resource recovery and circular-economy objectives. In this study, polyethersulfone (PES) mixed-matrix membranes incorporated with graphene oxide-titanium dioxide (GO-TiO2) nanocomposites were engineered to enhance photocatalytic degradation and minimize fouling under real wastewater conditions. The best membrane (PM1, 0.05 wt% GO-TiO2) exhibited reduced hydrophobicity, enhanced mechanical stability, and improved selectivity, achieving 98% BSA rejection and up to 85.4% COD removal using a low UV input (24 W). Reusability tests with the PM1 membrane demonstrated strong resistance to fouling, with flux recovery ratios in the range 81%-85%. To enable proactive fouling management, machine learning models were trained using 755 experimental records and validated against 1202 additional observations. Among the algorithms tested, Support Vector Regression (SVR) achieved the highest predictive accuracy (R-2 = 0.995, RMSE approximate to 0.30, MAE approximate to 0.24), outperforming Ridge, Lasso, and ElasticNet, particularly under high flux saturation conditions (R-2 > 0.99). These results emphasize the efficacy of the integrated membrane-ML approach for fouling detection, accurate flux prediction, and enhanced process control. With high removal efficiencies for key contaminants and nutrients (NH4+: 98%, PO43-: 96%), this strategy presents an energy-efficient, scalable solution for industrial wastewater treatment, supporting circular economy goals in pharmaceutical industry.