Machine learning enables interpretable discovery of innovative polymers for gas separation membranes

Area of Science:

• Materials Science
• Chemical Engineering
• Computational Chemistry

Background:

• Polymer membrane development for separations relies heavily on traditional, time-consuming methods.
• There is a critical need for accelerated discovery of novel polymer membrane materials with enhanced performance for environmental applications.

Purpose of the Study:

• To create a generalizable and accurate machine learning (ML) model for predicting polymer properties.
• To accelerate the discovery of innovative polymers for gas separation membranes.

Main Methods:

• Trained multitask ML models on experimental data linking polymer chemistry to gas permeabilities (He, H2, O2, N2, CO2, CH4).
• Interpreted ML models to understand chemical moiety contributions to permeability and selectivity.
• Screened over 9 million hypothetical polymers using the ML models.

Main Results:

• Identified thousands of polymers exceeding current performance upper bounds.
• Discovered hundreds of novel ultrapermeable polymer membranes with exceptionally high O2 and CO2 permeability.
• ML-predicted properties of top candidates were validated by molecular dynamics simulations.

Conclusions:

• The ML approach offers a powerful and accurate method for designing high-performance polymer membranes.
• This strategy can significantly expedite the development of advanced materials for gas separations.
• The identified polymers show potential for real-world translation in environmental applications.