Discovery of metal–organic frameworks for inverse CO2/C2H2 separation by synergizing molecular simulation and machine learning

Abstract

Separation of carbon dioxide (CO₂) from acetylene (C₂H₂) represents a significant challenge in the petrochemical industry, primarily due to their similar physicochemical properties. By synergizing molecular simulation (MS) and machine learning (ML), in this study, we aim to discover top-performing metal–organic frameworks (MOFs) for inverse CO₂/C₂H₂ separation. Initially, the adsorption of a CO₂/C₂H₂ mixture in MOFs from the Cambridge Structural Database (CSD) is evaluated through MS, structure–performance relationships are constructed, and top-performing CSD MOFs are shortlisted. Subsequently, ML models are trained by utilizing pore geometry, framework chemistry, as well as adsorption heat and Henry's constant as descriptors. The significance of these descriptors is quantitatively assessed through Gini impurity measures and Shapley additive explanations. Finally, the transferability of the ML models is evaluated through out-of-sample predictions for CO₂/C₂H₂ separation in the computation-ready experimental (CoRE) MOFs. Notably, a handful of CoRE MOFs are found to outperform the best CSD MOFs and their performance is further compared with existing literature. The synergized MS and ML approach in this study is anticipated to accelerate the discovery of MOFs in a large chemical space for CO₂/C₂H₂ separation and other important separation processes.