Two new studies demonstrate the potential of machine learning to accelerate the discovery of metal-organic frameworks (MOFs) that help capture carbon dioxide from the ambient air. First, researchers from the UK and South Korea screened 8,000 candidates MOFS and proposed new candidates for direct air capture (DAC).1 The second is a Prilint from the US, where researchers from Meta and elsewhere have released machine learning algorithms trained at 15,000 MOFs. 2
As a result of the ever-growing carbon emissions around the world, the Earth is increasingly likely to require the capture of carbon dioxide directly from the air to stabilize its climate. MOF is a candidate to lock it up. Unfortunately, they tend to preferentially absorb other gases, especially their strong dipole moments. “If you have to remove all the water from the sky, you've given up more joints2 Along the way, says Andrew Medford, a chemical engineer at Georgia Tech in Atlanta. Therefore, MOFs are required to selectively adsorb carbon dioxide in the presence of other gases.
Researchers can predict the adsorption properties of MOFS using density functional theory (DFT), but the computational cost of this technique severely limits the vast number of candidates that can be searched. A simpler method involving a “force field” similar to the ball-and-stick model of the MOF cage construction was originally used instead. As these proved to be ineffective in finding the adsorbent MOF selectively, Medford and colleagues were able to work with META researchers to perform multiple DFT simulations of MOF structures, reshaping their shape in response to guest molecules, and then develop a mechanical working algorithm that could predict the adsorbent properties of MOFs that do not require complete DFT calculations. The results, released as Open Direct Air Capture 23 (ODAC23), found some material considered Bind Co.2 More selective than water,” says Medford.
Several research groups have expanded the improved field of forces developed to search for a vast array of possible structures using machine learning, and researchers from Imperial College London and the Korea Institute of Advanced Science and Technology are currently presenting the results of searches for 8,000 candidate MOFs. They wanted not only potentially promising structures, but also the potential energy landscapes associated with them. Aron Walsh of Imperial College says this is “useful when you want to know what happens when multiple molecules are interacting, or when the pressure is too high and the system breaks. As a result, Walsh has identified several MOFs that were actually considered useless before, deserving further investigation.
In a future study based on 70 million new DFT calculations, Medford will present the ODAC25 dataset and its accompanying machine learning ability fields, along with colleagues at Oak Ridge National Laboratory in Meta, Tennessee and other countries. This provides information on variable adsorption properties of 15,000 MOF, providing competitive adsorption of nitrogen and oxygen, as well as water and water. We also investigate the effects of structural defects and amine functionalization at some sites. The results are currently unknown. Medford says researchers hope that the data will ultimately allow chemists to design and optimize MOFs for specific applications. “If you want to do direct air capture in Texas, where there is a high humidity level, you may need a different material than if you want to do direct air capture in Utah, where there is a low humidity level,” he says.
Walsh says he hopes that other types of artificial intelligence will help them find new MOFs in the future, rather than simply screening existing ones. “The landscape with the potential for different chemical constituents is so big that you need to be very smart about how to choose the right building block,” he says. “It leads to generative AI, reinforcement learning, and active learning. I think that's where the community really tries to design with the MOFS of the future, so that it's a place where the community is moving next to it.”
Shyue Ping Ong, a computational materials scientist at the University of California, San Diego believes that ODAC25 in particular is making an important contribution to this field. “This is an open source dataset available to all researchers, and by tweaking this open DAC dataset, we can better predict certain things, such as the binding energy of MOFS. He explains that the UK/Korean work is “somewhat uninteresting” as it is based on showing previously invisible possibilities of MOF. “When you claim that you have discovered new material, you want to see verification from the experimenter,” he says.
