Large-Scale Computational Screening of Metal-Organic Framework Membranes for Ethane/Ethylene Separation

Compared to the traditional heat-driven cryogenic distillation process, the membrane separation based on metal-organic frameworks (MOFs) is a technically and economically viable alternative for ethane/ethylene (C2H6/C2H4) separation. To accelerate the application of MOF membranes in this gas separation field, this study performed a large-scale computational screening of 12,020 real MOFs for the identification of optimal C2H6-selective MOF membrane materials. According to geometric and chemical analyses, 2,192 MOFs without open metal sites and having pore limiting diameter no less than 0.38 nm were first screened out. Grand canonical Monte Carlo and molecular dynamics simulations were subsequently carried out to mimic the adsorption and diffusion behaviors of ethane and ethylene in these MOFs respectively, based on which their C2H6/C2H4 membrane selectivities and C2H6 permeabilities were estimated. The results showed that MISQIQ04 exhibited the highest C2H6/C2H4 membrane selectivity (4.16) and moderate C2H6 permeability (4.35x10(5) Barrer). Additionally, structure-performance relationships between the C2H6/C2H4 membrane selectivity and structural properties of MOFs were investigated, covering the largest cavity diameter (LCD), pore limiting diameter (PLD), density (rho), gravimetric surface area (GSA), void fraction (VF), and pore volume (PV). The results indicated that MOFs with the structural characteristics of 0.4 nm <= LCD <= 1 nm, 0.38 nm <= PLD <= 0.75 nm, 0.8 g/cm(3)<= rho <= 2.5 g/cm(3), GSA=1,700 m(2)/g, 0.3 <= VF <= 0.73, and PV <=.85 cm(3)/g are optimal membrane materials for C2H6/C2H4 separation. Finally, a machine learning (ML) classifier was developed to achieve rapid prescreening of high-performing MOF membranes from a large MOF database, whose transferability was discussed on a hypothetical MOF database. Further t-Distributed Stochastic Neighbor Embedding analysis revealed that the ML model developed merely relying on a single MOF dataset generally exhibited poor transferability. Selecting the most representative and diverse MOFs from the entire MOF space for model development can help to improve the transferability and generalization ability of the developed model.