MOF-based type III porous liquids, comprising porous MOFs dissolved in a liquid solvent, have attracted increasing attention in carbon capture. However, discovering appropriate MOFs to prepare porous liquids was still limited in experiments, wasting time and energy. In this study, we have used the density functional theory and molecular dynamics simulation methods to identify 4530 MOF candidates as the core database based on the idea of prohibiting the pore occupancy of porous liquids by the solvent, [DBU-PEG][NTf2] ionic liquid. Based on high-throughput molecular simulation, random forest machine learning models were first trained to predict the CO2 sorption and the CO2/N2 sorption selectivity of MOFs to screen the MOFs to prepare porous liquids. The feature importance was inferred based on Shapley Additive Explanations (SHAP) interpretation, and the ranking of the top 5 descriptors for sorption/selectivity trade-off (TSN) was gravimetric surface area (GSA) > porosity > density > metal fraction > pore size distribution (PSD, 3.5-4 Å). RICBEM was predicted to be one candidate for preparing porous liquid with CO2 sorption capacity of 20.87 mmol/g and CO2/N2 sorption selectivity of 16.75. The experimental results showed that the RICBEM-based porous liquid was successfully synthesized with CO2 sorption capacity of 2.21 mmol/g and CO2/N2 sorption selectivity of 63.2, the best carbon capture performance known to date. Such a screening method would advance the screening of cores and solvents for preparing type III porous liquids with different applications by addressing corresponding factors.
Keywords: carbon capture; high-throughput molecular simulation; machine learning (ML); metal organic framework (MOF); porous liquid (PL).