Optimized prediction modeling of micropollutant removal efficiency in forward osmosis membrane systems using explainable machine learning algorithms

This study investigated the feasibility of using machine learning (ML)-based models to simulate the behavior of micropollutants (MPs) in the forward osmosis (FO) membrane water treatment process. Support vector regression (SVR) was combined with three optimization algorithms: firefly algorithm (FFA), particle swarm optimization (PSO), and grey wolf optimization (GWO) to establish the hybrid models. This study considered the rejection rates of 97 MPs using the osmotic membrane wastewater system. The ML models were constructed using nine input variables, which encompassed membrane characteristics, MPs properties, and experimental conditions. The developed models exhibited good prediction performance for simulating the MPs behavior in the FO process. The SVR-FFA model was found to be a better choice for simulating the MPs behavior in the FO process and exhibited higher accuracy with an R-value of 0.970 in testing and 0.969 in training. Furthermore, the SVR-FFA yielded predictions with mean absolute error (MAE) value of 2.353 in testing and 2.274 in training. The molecular weight of MPs exhibited the highest mean SHapley Additive exPlanation (SHAP) value, indicating its importance in influencing the MPs behavior and rejection in the FO process. The SVR-FFA model, coupled with SHAP interpretability, proved to be effective in predicting MPs rejection in the FO process. This contribution significantly enhances system design and operational efficiency, paving the way for achieving greater elimination of each MP in the future.