Discovering metal-organic framework/polymer mixed-matrix membranes via machine learning for CO2 separation

Mixed-matrix membranes (MMMs) that incorporate metal-organic frameworks (MOFs) as fillers surpass the permeability-selectivity trade-off of polymeric membranes. However, the enormous chemical diversity of MOFs and high computational cost of molecular simulations have hindered the systematic evaluation of large numbers of MOF/polymer MMMs. In this work, we present a data-driven discovery framework that integrates molecular simulations and machine learning (ML) to predict the CO2, CH4, N2, and H2 permeabilities of 104,196 different types of MOF/polymer MMMs. Molecular simulations were performed to obtain gas adsorption and diffusion properties for 3,982 synthesized MOFs and 4,701 hypothetical MOFs. Leveraging this dataset, we developed three ML approaches to enable rapid prediction of MMMs’ permeabilities and benchmarked their performance against both simulation results and experimental measurements. Our results revealed that many MOF/polymer MMMs exceed the upper bounds by achieving very high CO2 permeabilities and MOFs’ structural and chemical features play a decisive role in fine-tuning MMM performance. Mixed matrix membranes with metal organic frameworks can overcome the permeability selectivity trade off, but their diversity and simulation cost limit large scale evaluation. Here, the authors present a data driven framework using simulations and machine learning to predict gas permeabilities of 104,196 metal organic framework polymer membranes.