Interpretable machine learning and Magpie-based featurization for bridging water flux and physiochemical transport in forward osmosis membranes

Addressing global freshwater scarcity necessitates the development of high-performance separation and purification technologies, among which forward osmosis (FO) has emerged as a promising candidate. However, the advancement of robust models for Water Flux prediction in FO membranes is currently hindered by limited physiochemical diversity in training datasets and oversimplified feature engineering. In this study, we propose an advanced informatics framework that integrates the Materials Agnostic Platform for Informatics and Exploration (Magpie) featurization scheme to capture high-dimensional descriptors of draw solute (DS) physiochemistry, while accounting for systemic membrane configurations via discrete design indicators. By augmenting a compiled dataset with previously underrepresented DS types, the developed framework allows to achieve superior predictive fidelity compared to conventional approaches, while maintaining excellent model generalizability without overfitting. Beyond mere prediction, the framework employs also a multi-scale interpretability approach: Shapley Additive exPlanations (SHAP) analysis identifies Membrane Type, DS Molarity, Feed Solution (FS) Molarity, and Membrane Orientation as the primary Water Flux determinants, while a combined Partial Dependence Plot (PDP) – Individual Conditional Expectation (ICE) analysis delineates the non-linear, synergistic thresholds of structural and operational parameters of FO membranes. The presented study provides a practical and interpretable, data-driven, hierarchical informatics framework for prediction and interpretation of Water Flux in FO membranes, offering physiochemically-consistent interpretations of key transport-related factors, including concentration polarization and thermodynamic driving-force coupling, and serving as a quantitative guidance for analysis and optimization of membrane performance.