Machine learning accelerates high throughput design and screening of MOF mixed-matrix membranes towards He separation

Helium (He) is a non-renewable resource and membrane technology shows great potential for separating He from natural gas. Mixed matrix membrane (MMM) is a promising candidate for He separation, while the design of MMMs is still in the stage of “trial and error” due to the wide variety of fillers and complex membrane structures. In this work, we report machine learning accelerated high throughput design and screening of mixed matrix membranes. High throughput grand canonical Monte Carlo calculations and molecular dynamics simulations were performed to calculate the reliable structure and performance features of thousands of MOFs fillers. The high throughput computational data were combined with experimental data of polymer membranes to construct a big dataset of 456,872 MMMs. Four machine learning regression models were trained to predict the permeation and the separation performances and to reveal the underlying structure-performance relationships of MMMs. Our studies show that the XGBoost model has the best performance indicated by high R² value and lower RMSE and MAE values compared with the other models. The thorough data-driven model interpretation quantitatively revealed the key properties and structural parameters for fabricating MMMs with high He/CH₄ separation performance: for polymer matrices, S¹_He/CH4 > 900, P¹_He > 1600 Barrer and P¹_CH4 as low as possible; meanwhile for MOF fillers, porosity (ɸ) > 0.5 and 2.6 Å < PLD <3.2 Å. The XGBoost model shows good transferability on the dataset constructed from the hypothetical MOFs. This work provides guidance for designing He separation membranes.