Researchers at Koç University have developed a computational framework that uses machine learning to evaluate more than 100,000 mixed-matrix membrane combinations for gas separation. The study, led by master's student Feride Neva Yüngül and Professor Seda Keskin, was published in Communications Materials.
Mixed-matrix membranes embed metal-organic frameworks (MOFs) within polymer matrices to overcome a longstanding trade-off between permeability and selectivity in polymer-only membranes. With over 150,000 known MOF structures, identifying optimal pairings with commercially relevant polymers has been a major bottleneck.
The team paired 8,683 experimentally synthesized and computationally generated MOFs with 12 commercial polymers, creating a dataset of 104,196 combinations. Molecular simulations calculated gas interactions for carbon dioxide, methane, nitrogen, and hydrogen within each MOF, and these results trained three machine-learning models.
The most accurate model used a two-step strategy: it first predicted the gas permeability of the MOF alone, then combined that with polymer properties to estimate the full membrane performance. This approach achieved typical prediction errors of 10%–15% while requiring less computational effort than more complex alternatives.
The framework evaluated membranes for three industrial separations: carbon dioxide from methane for natural gas purification, carbon dioxide from nitrogen for post-combustion carbon capture, and hydrogen from carbon dioxide for hydrogen purification. Many predicted membranes surpassed the performance limits of conventional polymers, and some hydrogen-purification candidates improved both permeability and selectivity simultaneously.
MOF pore size emerged as a key factor controlling performance. Larger pores generally enabled faster gas transport but could reduce selectivity, while appropriately sized pores improved separation of specific gas mixtures. The model also provides guidance on which MOF characteristics enhance a given polymer and which combinations are unlikely to perform well.
Experimental production and characterization of a single mixed-matrix membrane can take months. The new framework estimates performance within seconds, serving as a rapid screening tool to prioritize the most promising candidates before laboratory synthesis. The researchers have made their predictive models and datasets publicly available to enable wider use by groups without extensive computational infrastructure.
AI screens 100,000+ membrane combinations, predicting carbon capture performance within seconds
This is an independent summary. The complete reporting, supporting context and any primary documents remain with Phys.org.