Predicting and screening high-performance polyimide membranes using negative correlation based deep ensemble methods.

Area of Science:

• Materials Science
• Computational Chemistry
• Chemical Engineering

Background:

• Polyimide membranes are vital for gas separation but exploring their vast chemical space experimentally is challenging.
• Over 10^7 polyimide types exist, making it difficult to identify optimal candidates for high selectivity and permeability.
• Current methods risk overlooking high-performance polyimide materials for gas separation and storage.

Purpose of the Study:

• To develop a deep learning model for rapid and efficient screening of polyimide structures for gas separation applications.
• To predict gas permeability and selectivity of polyimide materials, enabling faster discovery of novel candidates.
• To establish a link between polyimide chemical structure and gas separation performance.

Main Methods:

• A deep neural network model based on negative correlation learning (DNN-NCL) was developed.
• Morgan molecular fingerprints were used as input features for the deep learning model.
• SHAP analysis was employed for model interpretability and identification of key functional groups.

Main Results:

• The DNN-NCL model achieved an R^2 of ~0.95 on the test set, outperforming recent models by 4%.
• High-throughput screening of over 8 million hypothetical polyimides identified hundreds of promising candidates, including 14 exceeding the Robeson upper bound for CO2/N2 separation.
• SHAP analysis identified carbonyl, thiophene, and ester groups as crucial for gas permeability in polyimides.

Conclusions:

• Deep learning offers a powerful and efficient approach to explore the extensive chemical space of polyimides for gas separation.
• The developed model accurately predicts material performance and provides insights into structure-property relationships.
• This methodology accelerates the discovery of high-performance polyimide membranes for critical applications like CO2 capture.