Big Data Mining and Transferable Machine Learning for the High Accurate Prediction of Gas Separation Metal-Organic Framework Membranes

In this work, the physical properties of 21 gas binary mixtures were combined with the basic physical descriptors of metal-organic framework membranes (MOFMs) as descriptors to train the machine learning model for evaluating the gas permeation and separation performance of MOFMs. The highly accurate (R² = 0.935) extreme gradient boosting (XGBoost) model demonstrated outstanding generalization capabilities, effectively predicting the permeation performance of different gas binary mixtures in both the Computation-Ready Experimental MOFMs database and the hypothetical MOFMs database. Furthermore, the trained XGBoost model has been encapsulated to provide a user-friendly interface and allow other researchers to utilize the model for gas permeation predictions efficiently. The Shapley Additive Explanations analysis based on the trained XGBoost model was employed, revealing that gas polarizability is a key factor governing gas permeation performance. Additionally, two distinct membrane separation mechanisms dominated by adsorption and diffusion effects were identified for the CO₂/M, He/X and H₂/Y (M = CH₄, H₂S, N₂, O₂; X = CH₄, N₂, O₂, CO₂, H₂S, H₂; Y = CH₄, N₂, O₂, CO₂, H₂S) binary mixtures through big data mining. The data-driven approach enhances molecular permeation prediction, aiding the design of high-performance membranes for complex gas mixtures.