Two new studies have demonstrated the potential of machine learning to accelerate the discovery of metal–organic frameworks (MOFs) that could be useful for capturing carbon dioxide from ambient air. In the first, researchers in the UK and South Korea screened 8000 candidate MOFs and suggested some new candidates for direct air capture (DAC).1 In the second, a preprint from the US, researchers at Meta and elsewhere released a machine-learning algorithm trained on 15,000 MOFs. 2
As a result of the world’s ever-rising carbon emissions, the Earth is increasingly likely to require direct carbon dioxide capture from air to stabilise the climate. MOFs are promising candidates to trap it. Unfortunately, they tend to preferentially absorb other gases – especially water, thanks to its strong dipole moment. ‘If you have to take all the water out of the air, you’ve emitted more CO2 in the process,’ says chemical engineer Andrew Medford at Georgia Institute of Technology in Atlanta. A MOF that selectively adsorbs carbon dioxide in the presence of other gases is therefore needed.
Researchers can predict the adsorption properties of MOFs using density functional theory (DFT), but the technique’s computational cost severely restricts the huge number of candidates they can search. Simpler methods involving ‘force fields’ – analogous to ball-and-stick models of the MOF cage structures – were originally used instead. These proved effectively useless for finding selectively adsorbent MOFs, so Medford and colleagues teamed up with researchers at Meta to conduct multiple DFT simulations of the MOF structures, allowing them to change their shape in response to the guest molecules, and then to develop a machine-learning algorithm that could predict the adsorption properties of an unseen MOF without needing full DFT calculations. The results, which were released as Open Direct Air Capture 23 (ODAC23), ‘did find some materials deemed to bind CO2 more selectively than water’, says Medford.
Multiple research groups have extended improved force fields developed using machine learning to search huge arrays of possible structures, and researchers at Imperial College London and Korea Advanced Institute of Science and Technology now present the results of a search of 8000 candidate MOFs. They looked not just for potentially promising structures, but for the potential energy landscapes associated with them. Aron Walsh at Imperial College says this is useful ‘if you want to know what happens when you have multiple molecules interacting or what happens when you have too high a pressure and the system breaks down’. As a result, Walsh says, they identified several MOFs previously considered useless that are, in fact, worthy of further investigation.
In forthcoming research based on 70 million new DFT calculations, Medford, together with colleagues at Meta, Oak Ridge National Laboratory in Tennessee and elsewhere, unveil the ODAC25 dataset and accompanying machine-learning force fields. This gives information on the variable adsorption properties of 15,000 MOFs and studies the competitive adsorption of nitrogen and oxygen, as well as water. It also investigates the effects of structural defects and amine functionalisation of some sites. At present, the results are unclear. Medford says that the researchers hope the data will ultimately allow chemists to design and optimise MOFs for specific applications. ‘If I want to do direct air capture in Texas where the humidity is high, maybe I need a different material than if I want to do direct air capture in Utah, where the humidity is low,’ he says.
Walsh says that he hopes other sorts of artificial intelligence will, in future, help find new MOFs, as opposed to simply screening existing ones. ‘The possible landscape of different chemical components is so large, you need to be very smart with how you choose the right building blocks,’ he says. ‘That leads to generative AI, reinforcement learning, active learning… I think that’s where the community’s going next to really try to design in silico the MOFs of the future.’
Computational materials scientist Shyue Ping Ong at University of California, San Diego believes that the ODAC25 work in particular marks an important contribution to the field. ‘This is an open-source dataset that is available to all researchers, and they’ve shown that, by fine-tuning on this open DAC dataset you can get much better predictions of certain things like the binding energy for the MOFs – so that’s very interesting,’ he says. He describes the British/Korean work as ‘slightly less interesting’ because it is predicated on demonstrating previously unseen potential of MOFs. ‘When you make a claim that “I have discovered a new material” then preferentially I want to see validation from an experimentalist,’ he says.