In this work, the separation performance of methane/ethane/propane (C₁, C₂ and C₃) mixture in the 137953 hypothetical metal-organic frameworks (MOFs) is calculated by high throughput computational screening and multiple machine learning (ML) algorithms. First, to avoid the competitive adsorption of water vapor, 31399 hydrophobic MOFs (hMOFs) were screened out. Then, grand canonical Monte Carlo (GCMC) simulations were employed to calculate the adsorption behavior of a mixture with a mole ratio of C₁:C₂:C₃=7:2:1 in these hMOFs, respectively. Second, the relationships among six MOF structures/energy descriptors (the largest cavity diameter (LCD), void fraction (f), volumetric surface area (VSA), Henry coefficient (K), heat of adsorption (Q_st), density of MOF (ρ)) and three performance indicators of MOFs (selectivities (S), adsorption capacities (N) of C₁, C₂, C₃ and their trade-offs (TSN)) were established. The LCDs were calculated by Zeo++software, and VSAs were calculated using RASPA software using He and N₂ as probes, respectively, and Q_st and K were calculated in an infinite dilution of each gas molecule in an infinite dilution state using NVT-MC method in RASPA software. Then, we found that there existed the "second peaks" of N and S in part of structure-property relationships, and all the optimal MOFs located in the range of second peaks, especially for the separation of C₁ or C₂. Third, the above-mentioned six MOF descriptors and three MOF performance indicators were trained, tested and predicted by four ML algorithms, including decision tree, random forest (RF), support vector machine and Back Propagation neural network. Although the predictive effect for the selectivity was very low, the introduction of TSN can significantly improve the accuracy of ML prediction, especially for RF algorithm (R=0.99). Therefore, the RF was used to quantitatively analyze the relative importance of each MOF descriptor, and found that three descriptors (K, LCD and ρ) possessed the highest importance for the separation of C₁ and C₂, and three other descriptors (K, Q_st and ρ) for the separation of C₃. Moreover, three simple and clear paths of optimal MOFs for C₁, C₂ and C₃ adsorption were designed by the decision tree model with the descriptors. Based on those paths, there were 96%, 85%, 95% probability that we can search for high-performance MOFs, respectively. Finally, the best 18 MOFs were identified for different separation applications of C₁, C₂ and C₃. This study reveals the second peaks and key MOF descriptors governing the adsorption of light alkane, develops quantitative structure-property relationships by ML, and identifies the best adsorbents from a large collection of MOFs for the separation of C₁, C₂ and C₃ from natural gas.

Key words: metal-organic framework, gas separation, molecular simulation, machine learning