Optimizing water-salt separation performance and mechanistic analysis of TFNi nanofiltration membranes via explainable machine learning

Addressing global water scarcity requires the advancement of water treatment technologies that combine high efficiency with minimal energy consumption. Thin-film nanocomposite nanofiltration membranes featuring interlayer structures (TFNi) have garnered significant interest due to their exceptional capacity for water–salt separation. Nevertheless, the intricate interplay between membrane morphology and operational parameters complicates systematic performance optimization. In this work, we established a machine learning framework designed for both prediction and interpretation, focusing on water flux and salt rejection as key performance indicators. Four machine learning algorithms were evaluated, with the CatBoost model demonstrating superior predictive accuracy (R² > 0.90). Feature importance analysis using Shapley additive explanations identified applied pressure, temperature, and molar concentration as the predominant factors influencing water flux, while salt rejection was primarily governed by molar concentration, membrane pore size, and contact angle. Further analysis using partial dependence plots revealed clear nonlinear responses. For water flux prediction, higher response levels were identified near 27 °C, around 0.04 mol·L⁻¹, and within the medium-to-high pressure range. For salt rejection prediction, a moderate pore size range (approximately 40–80 nm) maintained relatively high rejection levels when combined with surface wettability. By coupling data-driven prediction with interpretable artificial intelligence techniques, this study delineates the influence and tendencies of critical variables, offering a principled strategy for guiding the rational design and operational optimization of TFNi nanofiltration membranes.