Machine learning-assisted discovery of high-performance polymer membranes for gas separation

High-performance polymer membranes have achieved remarkable success in gas separations. Contrary to the traditional Edisonian trial-and-error approach, the growing machine learning (ML) technique possesses great promise to accelerate the discovery and development of innovative polymer membrane materials yet is obstructed by the insufficiency and label imbalance of training data. We demonstrate the success of a state-of-the-art semi-supervised graph regression framework leveraging unlabeled polymers to improve the performance of ML models even if only very limited and imbalanced permeability data are available based on an open-source polymer gas permeability experimental database for six major industrial gases. Using the trained models, gas permeability prediction is conducted on over 12,000 existing unlabeled homopolymer candidates and the prediction accuracy is tested by experimentally synthesizing two of the most promising polymer membranes, which are found to possess extraordinary H₂/CH₄, H₂/N₂, and O₂/N₂ separation performance, representing an advanced way to explore the unknown chemical space for high-performance gas separation polymer membrane design.