Data-Driven Prospective Design of Phase-Inversion Conductive Polymer Membranes

Abstract

Conductive polymer membranes combine separation and electrical functionality, but their design remains challenging because performance emerges from coupled effects of composition, fabrication, and operating conditions. Here, we assembled 396 membrane records from 77 publications and developed target-specific machine-learning models for pure-water permeability (PWP), solute rejection, and electrical conductivity. Weighted artificial neural network ensembles for PWP and rejection and XGBoost for conductivity achieved fixed-holdout R2 values of 0.855, 0.896, and 0.914, respectively, with distinct SHAP feature-contribution patterns across the three targets. We then coupled the predictive models with residual Gaussian-process uncertainty estimation and sequential Monte Carlo expected hypervolume improvement to search a 3,024-member PANI–DBSA/PES design space and identify 12 formulations for prospective experimental evaluation. Atomistic molecular dynamics simulations provided complementary mechanistic insight: increasing PANI from 0.47 to 1.89 wt% decreased the first-shell hydration number around DBSA sulfonate sites from 1.02 to 0.74, while the apparent finite-time water diffusion coefficient declined from 0.6910-10 m2∙s-1 in neat PES to 0.3710-10 m2∙s-1 at 1.89 wt% PANI. Together, these results connect literature-derived learning, uncertainty-aware formulation selection, and molecular-scale interpretation, providing a systematic framework for prospective conductive membrane design