Machine learning-based prediction and optimization of polymeric membranes for CO2 separation

Abstract

Polymeric membranes are widely used for gas separation due to their energy efficiency and scalability, particularly for carbon dioxide (CO2) capture applications. However, accurately predicting gas permeability in polymeric membranes remains a challenge due to complex nonlinear structure–property relationships and the limited availability of experimental data. To overcome this challenge, this work proposes a machine learning (ML) framework that incorporates uncertainty estimation. Three ML models based on ensemble trees methods, such as random forest, extreme gradient boosting (XGBoost), and CatBoost, were used to predict the gas permeability. Furthermore, Extended Connectivity Fingerprints were used to represent the structural features of polymers. In addition, hyperparameter optimization with five-fold cross-validation was performed to improve the predictive performance and robustness of the models. Among the evaluated models, the optimized XGBoost model achieved the best performance for CO2 permeability prediction, with a coefficient of determination of 0.9410, mean absolute error of 0.0711, root mean square error of 0.1425, and mean absolute scaled error of 0.1147. Furthermore, model interpretability was analyzed using feature importance and SHAP methods to understand the relationship between polymeric structure and permeability. Moreover, uncertainty quantification was incorporated to improve prediction reliability in the virtual screening of polymeric membranes. The proposed framework further enables permeability–selectivity trade-off analysis, Robeson upper bound comparison, Pareto-optimal candidate identification, and screening-level feasibility analysis for identifying promising membrane materials.