Machine Learning Prediction of Henry Coefficients of Polar and Nonpolar Gases in Covalent Organic Frameworks: Effects of Interlayer Shifts and Functionalization

Covalent organic frameworks (COFs) are promising materials for gas separation and carbon capture, however, the vast chemical and structural space of possible COFs makes conventional molecular simulation‐based screening of adsorption properties such as Henry coefficients computationally infeasible. Here, we systematically investigate the performance of several physically motivated descriptors, the sine matrix, Ewald sum matrix, smooth overlap of atomic positions (SOAP), many‐body tensor representation (MBTR), and a custom Lennard Jones‐based descriptor for predicting Henry coefficients of small gas molecules (both polar and nonpolar) in COFs using machine learning. To account for realistic variability, we construct datasets including both chemically functionalized frameworks and interlayer displaced stacking configurations, and train relatively simple models (random forests, neural networks, and gaussian process regression) on these data. By comparing predictive performance across descriptor–model combinations, we demonstrate how various physical representations capture the key factors governing gas adsorption, including electrostatics, local atomic environments, and van der Waals interactions. Our results highlight the critical role of descriptor choice and provide physically interpretable guidance for designing efficient machine learning models, providing a foundation for scalable high throughput computational screening of COFs for gas separation applications.