Module-based machine learning models using sigma profiles of organic linkers to predict gaseous adsorption in metal-organic frameworks

Background: Metal-organic frameworks (MOFs) have drawn considerable attention for their potential in adsorption applications, such as gas separation and storage. Machine learning (ML) augmented high-throughput screening approaches have emerged as an effective strategy to expedite the materials search. Traditionally, ML models developed to predict the adsorption properties of MOFs rely on various geometrical and chemical descriptors. While these descriptors are effective, they tend to be specific to each MOF's unique structure, completely omitting the modular nature of MOFs. Methods: A new approach is proposed in this study: a modular descriptor based on the sigma profile of MOF organic linkers. These sigma profiles effectively represent the chemical environment of organic linkers. With these profiles as input features, we train extreme gradient boosting (XGBoost) models to predict the Henry's coefficient (KH) of adsorption for hydrocarbons and acid gases in MOFs. Findings: The results show that sigma profiles enhance the prediction accuracy and emerge as the most important features for hydrocarbon gases. This study highlights the potential of sigma profiles in developing accurate ML models for identifying optimal MOF adsorbents. Such an approach could also facilitate an inverse design of MOFs with targeted properties.