Interpretable Machine‐Learning and Big Data Mining to Predict the CO2 Separation in Polymer‐MOF Mixed Matrix Membranes

Abstract Mixed matrix membranes (MMMs) are renowned for their exceptional gas separation capabilities. In this work, high‐throughput computing screening and machine learning are employed to evaluate the CO 2 separation performance of 54117 MMMs composed of 9 polymers and 6013 metal–organic frameworks (MOFs). The structure‐property relationships of MMMs are analyzed for 4 binary mixtures (CO 2 /X, X = CH 4 , N 2 , H 2 , O 2 ), and the best‐performing combinations of MOFs and polymers are found, with which the MMM performance exceeded the Robeson's upper limit. Then, a stacked ensemble regression model with high accuracy (average R 2 = 0.96) is trained, demonstrating excellent extrapolation capability ( R 2 = 0.95) for new MMMs containing 6FDA‐DAM. Furthermore, by utilizing Shapley Additive Explanations and data segmentation, it is identified that the pore limit diameter and largest cavity diameter in MOF features and the fractional free volume and density in polymer features are of paramount importance. Two extrapolation methods are compared and found that transfer learning is better for predicting CO 2 separation performance in MMMs and designing new materials with large datasets. Finally, an interactive desktop software is developed to assist researchers in rapidly and accurately calculating the CO 2 separation performance of MMMs. This work presents a novel approach for the rapid evaluation of high‐quality MMMs and the efficient calculation of gas permeation rates within membranes.