Machine learning-based response and prediction analysis for gravity-driven membrane

Gravity-driven membrane (GDM) technology has emerged as a promising approach for wastewater treatment due to its low energy demand and sustained operational performance. In this study, the predictive model was constructed to clarify the interactions among membrane parameters, operational parameters, and feed water quality parameters by machine learning (ML) methods combining recursive feature elimination (RFE) and extreme gradient boosting (XGBoost). A refined dataset of 515 samples and 11 features was created through RFE-based feature selection, ensuring robust model performance and a high signal-to-noise ratio. The results showed that the XGBoost model demonstrated superior predictive accuracy (R² = 0.93, RMSE = 3.55) than RFE-random forest (RFE-RF), RFE-decision tree (RFE-DT), and RFE-multiple linear regression (RFE-MLR). Shapely Additive Explanations (SHAP) analysis, based on the optimized RFE-XGBoost, revealed that ceramic membrane, operating day, and pretreatments were the most influential parameters in predicting flux. The SHAP analysis further revealed that ceramic membranes played a dominant role in increasing flux, whereas membrane pore size exhibited a threshold effect, requiring a trade-off between permeability and fouling resistance. Additional findings indicated that short-term operation (<25 days) and optimized pretreatment effectively mitigated flux decline. The synergistic control pH (7.4–7.8), influent chemical oxygen demand (COD)(<4 mg/L), and influent dissolved oxygen (DO)(>7.3 mg/L) increased flux through electrostatic repulsion and aerobic degradation mechanisms. Additionally, a graphical user interface (GUI) was developed for real-time flux prediction and parameter optimization, validated with 75 data points from four papers (2020–2024), with most flux prediction errors within 20 %. This work could provide theoretical support for improving GDM system performance and advancing its practical application.