Interpretable machine learning for materials discovery: Predicting CO2 adsorption properties of metal–organic frameworks

Abstract

Metal–organic frameworks (MOFs), as novel porous crystalline materials with high porosity and a large specific surface area, have been increasingly utilized for CO2 adsorption. Machine learning (ML) combined with molecular simulations is used to identify MOFs with high CO2 adsorption capacity from millions of MOF structures. In this study, 23 structural and molecular features and 765 calculated features were proposed for the ML model and trained on a hypothetical MOF dataset for CO2 adsorption at different pressures. The calculated features improved the prediction accuracy of the ML model by 15%–20% and revealed its interpretability, consistent with the analysis of the interaction potential. Subsequently, the importance of the relevant features was ranked at different pressures. Regardless of the pressure, the molecular structure and pore size were the most critical factors. van der Waals force-related descriptors gained more competitive advantages at low pressures, whereas electrical-field-related descriptors gradually dominated at high pressures. Overall, this study provides a novel perspective to guide the initial high-throughput screening of MOFs as high-performance CO2 adsorption materials.