Optimizing Ternary Preparation Parameters of Thin-Film Nanocomposite Membranes for Tailored Selectivity via Interpretable Machine Learning

Novel thin-film nanocomposite (TFN) nanofiltration membranes have attracted wide attention for their generally boosted water permeability. However, the incorporation of diverse nanomaterials introduces significant uncertainty in tuning the pore size of TFN membranes, thereby complicating the regulation of solute selectivity. This challenge is further amplified by the complex interactions among the substrate, loaded nanomaterials, and interfacial polymerization (IP) for the polyamide layer formation, hindering the efficient optimization of membrane fabrication parameters. In this study, the ternary preparation parameters (substrate, nanomaterials, and IP) were decoupled to tailor the pore-size-dominated selectivity of TFN membranes, with the molecular weight cutoff (MWCO) used as a key indicator predicted via interpretable machine learning. Critical factors for attaining a low membrane MWCO were revealed to include selecting a moderately dense substrate, optimizing nanomaterial load and monomer concentration, and ensuring compatibility between nanomaterial stacking and other preparation parameters. Furthermore, a novel SHAP fusion strategy was proposed to refine MWCO optimization by leveraging MgCl₂ rejection insights as performance constraints, aiming for simultaneously low MWCO (generally high removal of organics) and high mineral ion passage, which were usually desired in drinking water purification. This study provides a systematic understanding of TFN membrane formation and offers an efficient, data-driven approach for customizing nanofiltration membranes for advanced water treatment applications.