Practical Prediction of Gas Separation Performance in Polymeric Membranes Using Machine Learning

Membrane-based gas separation offers significant potential for carbon capture and clean energy applications; however, inconsistent experimental conditions often hinder direct comparison of membrane performance. In this work, machine learning (ML) is applied to predict gas permeability in polymeric membranes using a dataset comprising 3618 entries from 603 polymers across six gases (CO₂, N₂, H₂, He, O₂, and CH₄). Ensemble regression algorithms (Random Forest, Gradient Boosting, XGBoost, and Extra Trees) were developed and rigorously assessed. Model accuracy was evaluated using mean absolute error (MAE) and root mean square error (RMSE), both expressed in Barrer for physical relevance. Among the models, Random Forest exhibited the most reliable performance, achieving an MAE of 346.92 Barrer and an RMSE of 888.69 Barrer, highlighting its strong predictive capacity across diverse membrane-gas systems. Feature importance analysis, combined with SHAP interpretation, indicated that membrane structure, operating parameters, and gas-specific properties play key roles in determining permeability. Benchmarking predictions against the Robeson upper bound further identified high-performing membranes for CO₂/N₂ and CO₂/CH₂ separations. Overall, the study demonstrates the effectiveness of ML-driven screening in accelerating membrane discovery and supporting the design of next-generation gas separation materials.