Cambridge, UK-based computational chemistry company CuspAI has formed a network of 45 organisations, aimed at discovering new functional materials using artificial intelligence (AI). The AI Materials Foundry includes tech giants like Facebook parent company Meta and chipmaker Nvidia. It also involves high-profile companies spanning the automotive, chemical, solar power and semiconductor industries, as well as scientific data providers. CuspAI plans to work with the partners to make and test predictions for innovative new materials.
Predicting whether a material has a desired property is the easy part. The constraint has shifted to synthesis, stability under real operating conditions and manufacturability. My expectation is that value will accrue to those who achieve real-world impact.
CuspAI chief scientific officer and Imperial College London materials scientist Aron Walsh tells Chemistry World that working for the startup is interesting because of its ability to scale. ‘You can start to train bigger models with bigger datasets than typically would be accessible in an academic lab,’ Walsh says. CuspAI’s partnership with Meta and Nvidia, enables ‘access to capabilities and scale of resource’ for computing beyond most startups.
In 2025, CuspAI worked with Meta to develop datasets and AI models for materials that might capture carbon dioxide directly from the atmosphere. Walsh explains that the project was less about the business case for direct air capture materials than applying CuspAI’s techniques to ‘a design space that was quite rich’. The company’s AI methods ‘generate novel structures and amplify the structural and chemical diversity of the dataset, making it more useful to researchers and opening up new paths for discovery and impact,’ says CuspAI chief executive Chad Edwards.
Larger datasets enable more general models
As part of its ‘AI for materials platform’ CuspAI uses Meta’s Universal Model for Atoms (UMA), also released in 2025. UMA expands the applicability of machine learning interatomic potentials (MLIPs), a highly successful deployment of AI in chemistry. MLIPs can calculate atomic forces using less computational time and energy than density functional theory (DFT) approaches. Whereas scientists previously developed a new MLIP for each chemical system, UMA can generalise to different chemical systems.
Some of the underlying tools are available as open-source models. Others are proprietary, especially those trained on privately held data from partners. Walsh says that such models have progressively developed from often making chemically incorrect suggestions to now ‘getting very, very good’ at suggesting substances with specific property sets. ‘It’s this type of inverse design problem that traditionally has been very, very, difficult in chemistry and materials science,’ Walsh says.
One key driver behind this improvement has been developing sufficiently large datasets for training. To achieve this, CuspAI uses data from the Cambridge Crystallographic Data Centre and licences content from scientific publishers like Wiley, both of whom are AI Materials Foundry partners. CuspAI supplements such archives with simulated datasets with tens of millions of substances. It will also work with AI Materials Foundry partners to build its own experimental data sets, in particular A*Star, Singapore’s national lab.
Faster and more direct
Beyond its carbon capture work, an important case study for CuspAI comes from a collaboration with Finnish chemicals company Kemira. CuspAI suggested 20 novel metal–organic frameworks for removing per- and polyfluoroalkyl substance (PFAS) pollutants from water in six months rather than several years.
Walsh explains that together CuspAI’s tools avoid exhaustively searching infinite chemical space, instead learning to focus on promising possibilities. They use correlations between structural data and properties to do this, somewhat like a large language model assembles text based on correlations with surrounding words. He likens the learning to a chemist’s intuition for possible chemical compositions based on knowledge like oxidation states of specific elements. ‘You end up going from an infinite space to something that’s slightly more accessible,’ Walsh says. ‘For me, that’s the sort of the thing that’s quite exciting in AI for science.’
The science ‘gets a bit more serious or focused when you’re dealing with the direct industrial problems and processes’, Walsh says, hence the move to partner with more materials companies. ‘The materials that power a cleaner future won’t be discovered the way they used to be, which is why we’re proud to be a founding member of the AI Materials Foundry,’ said Elizabeth Rowsell, chief technology officer at UK-based catalyst producer Johnson Matthey on LinkedIn.
Alongside its partnership agreements, CuspAI has raised £330 million from investors, including Amazon’s founder Jeff Bezos and national funds from the UK and the Netherlands, where some CuspAI activities reside. The company earns revenue from its relationships in different ways, Walsh notes, but eventually wants to earn royalties from sales of materials it has helped discover.
Amid large amounts of investment in AI-related companies, there are concerns that the technology is overhyped. Walsh notes that he’s previously seen excitement around technologies such as high-throughput screening and combinatorial chemistry when they first emerged. ‘Each left something real behind, but each was oversold at the peak,’ he says.
What’s different today is that predicting whether a material has a desired property ‘is the easy part’, Walsh adds. ‘The constraint has shifted to synthesis, stability under real operating conditions and manufacturability,’ he says. ‘So, my expectation is that value will accrue to those who achieve real-world impact.’
No comments yet