Reducing carbon dioxide emissions from industrial processes and energy production remains one of the major technological challenges in addressing climate change. Membrane-based gas separation offers an energy-efficient alternative to conventional separation technologies, but identifying membranes that allow gases to pass through rapidly while also separating them effectively has long presented a major materials-design challenge.

Researchers at Koç University have now developed a data-driven framework that combines molecular simulations with machine learning to accelerate the discovery of high-performance membrane materials. The approach allowed the researchers to evaluate more than 100,000 material combinations and predict the performance of promising candidates within seconds.

The study, conducted by master’s student Feride Neva Yüngül and Prof. Dr. Seda Keskin from Koç University’s Department of Chemical and Biological Engineering, was published in Communications Materials, a Nature Portfolio journal.

Most commercial gas-separation membranes are made from polymers because they are inexpensive, easy to process and suitable for large-scale production. However, polymer membranes are constrained by a well-known trade-off between permeability and selectivity. Polymers that transport gases rapidly generally separate them less efficiently, while highly selective polymers tend to restrict gas transport.

One strategy for overcoming this limitation is to embed metal-organic frameworks, or MOFs, within polymer membranes. MOFs are porous crystalline materials constructed from metal ions and organic linkers. Because their pore sizes and chemical properties can be precisely tuned, they can selectively adsorb and transport particular gas molecules.

The resulting materials, known as mixed-matrix membranes, combine the scalability and processability of polymers with the gas-separation capabilities of MOFs. Yet selecting the right MOF-polymer pairing is difficult: more than 150,000 MOF structures have been reported, creating millions of potential combinations with existing polymers.

To explore this enormous design space, the researchers paired 8,683 experimentally synthesized and computationally generated MOFs with 12 commercially relevant polymers. This produced a dataset comprising 104,196 mixed-matrix membrane combinations.

Molecular simulations were used to calculate how carbon dioxide, methane, nitrogen and hydrogen interact with and move through each MOF. The resulting data were then used to train machine-learning models that could rapidly predict the permeability of previously unexplored MOF-polymer combinations.

Rather than relying on a single machine-learning method, the researchers developed and compared three predictive approaches using molecular simulation results and available experimental membrane data.

The most accurate model used a two-step strategy. It first predicted the gas permeability of the MOF and then combined this information with the properties of the polymer to estimate the performance of the mixed-matrix membrane. This approach produced more reliable predictions while requiring less computational effort than the more complex alternatives.

The researchers evaluated the membranes for three industrially important gas separations:

• Carbon dioxide from methane for natural gas purification
• Carbon dioxide from nitrogen for post-combustion carbon capture
• Hydrogen from carbon dioxide for hydrogen purification

The results indicated that many of the predicted mixed-matrix membranes could surpass the established permeability and selectivity limits of conventional polymer membranes.

Adding MOFs generally increased gas permeability. Some membranes evaluated for hydrogen purification were predicted to improve both permeability and selectivity simultaneously.

The study also identified MOF pore size as an important factor controlling membrane performance. Larger pores generally support faster gas transport but may reduce selectivity, while appropriately sized pores can improve the efficient separation of particular gas mixtures.

Beyond ranking promising candidates, the framework provides practical guidance about which MOF characteristics are most likely to enhance a particular polymer and which combinations are unlikely to perform well.

Producing and thoroughly characterizing a single mixed-matrix membrane experimentally can take months. The new framework can estimate the performance of a proposed material combination within seconds, with typical prediction errors of approximately 10–15%.

The researchers emphasize that the approach is intended to guide laboratory research, not replace it. Molecular simulations generally represent ideal crystalline MOFs, whereas experimentally produced membranes may contain structural defects and imperfections at the interfaces between MOFs and polymers.

The framework therefore serves as a rapid screening tool, enabling researchers to prioritize the most promising material combinations before committing substantial time and resources to synthesis and testing.

By making their predictive models and datasets publicly available, the researchers are also enabling groups without extensive computational infrastructure to evaluate thousands of potential membrane materials.

The approach could accelerate the development of more efficient and cost-effective membranes for carbon capture, natural gas purification and hydrogen production.