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

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Chemistry

Background:

• Membrane-based gas separation is crucial for carbon capture and clean energy.
• Inconsistent experimental data complicates direct comparison of membrane performance.

Purpose of the Study:

• To apply machine learning (ML) for predicting gas permeability in polymeric membranes.
• To develop and assess ensemble regression algorithms for enhanced predictive accuracy.
• To identify key factors influencing gas permeability and high-performing membranes.

Main Methods:

• Utilized a dataset of 3618 entries from 603 polymers across six gases (CO2, N2, H2, He, O2, CH4).
• Developed and evaluated ensemble regression algorithms: Random Forest, Gradient Boosting, XGBoost, and Extra Trees.
• Assessed model accuracy using Mean Absolute Error (MAE) and Root Mean Square Error (RMSE) in Barrer.
• Employed feature importance and SHAP interpretation for understanding predictive drivers.

Main Results:

• Random Forest demonstrated superior performance with MAE of 346.92 Barrer and RMSE of 888.69 Barrer.
• Membrane structure, operating conditions, and gas properties were identified as critical factors for permeability.
• ML predictions benchmarked against the Robeson upper bound identified promising membranes for CO2/N2 and CO2/CH2 separations.

Conclusions:

• Machine learning effectively predicts gas permeability in polymeric membranes.
• ML-driven screening accelerates the discovery and design of novel gas separation materials.
• This approach supports the development of next-generation materials for carbon capture and energy applications.