Learning the Path to Ion Separation: A Proof-of-Concept for Machine Learning-Assisted Inverse Design (MLAID) of Nanofiltration Membranes

Abstract

Abstract Advancing the design of polymeric membranes through material screening and discovery offers a powerful strategy to guide future membrane fabrication tailored to a wide range of separation challenges. Traditional trial-and-error methods often fall short in navigating this large monomer and process-design space, whereas machine learning-assisted inverse design (MLAID) can provide an extensible workflow for translating literature-derived membrane knowledge into ranked experimental candidates. Building on prior inverse-design strategies that combine machine learning (ML) with Bayesian optimization (BO), this study applies MLAID to prioritize membrane candidates within a target-defined nanofiltration (NF) design space. MLAID models are trained with membrane fabrication data, and their underlying patterns are interpreted through Shapley additive explanations (SHAP). SHAP was used to investigate model-attribution patterns, and BO was used to prioritize monomer and fabrication-condition combinations for producing polyamide membranes with targeted permeability and selectivity. The MLAID framework prioritized candidate formulations across target salt-water permeabilities of 4 to 12 LMH/bar (L m–2 h–1 bar–1), with Na2SO4 rejection used as the experimental benchmark. Fabrication and testing served as the experimental decision gate for evaluating the model-prioritized candidates. Overall, this study demonstrates how ML/BO-guided prioritization can accelerate identification of promising monomers and fabrication conditions within an encoded design space, while providing a workflow backbone that can expand as richer target-specific data sets, descriptors, and validation become available.