The development of high-performance polymeric membranes is critical for advancing energy-efficient CO 2 separation technologies. While machine learning (ML) has emerged as a powerful tool for predicting gas permeability and accelerating polymer discovery, most existing approaches rely on limited experimental descriptors and act as black boxes , offering little physical insight or reliability assessment. Here, we use ML as a molecular microscope to construct a physically informed and interpretable framework for predicting gas-transport properties in polymeric membranes for CO 2 capture. By integrating structural fingerprints, physicochemical descriptors, and operational variables such as temperature and pressure, the framework moves beyond the typical black-box paradigm toward physically interpretable structure-property relationships. The approach combines data imputation through Multiple Imputation by Chained Equations (MICE), regression modeling with ensemble algorithms, and rigorous applicability-domain analysis to ensure predictive reliability. Feature-importance and explainability analyses reveal the molecular signatures that control gas transport—rigid, branched, and moderately polar architectures that balance free volume and solubility. This interpretive capability transforms ML into an analytical tool for rational polymer design, bridging predictive accuracy with chemical understanding and providing quantitative guidance for the next generation of CO 2 -selective membranes. • Explainable Machine learning techniques used to uncover polymer transport behavior • Polymer structure strongly governs CO 2 permeability, revealed through data analysis • Missing gas-permeability values can be reliably reconstructed from existing data trends • Membrane permeability can be accurately predicted for a wide range of polymer structures • Insights provide practical structural guidelines for designing CO 2 - selective membranes
Vigna et al. (Sun,) studied this question.