Researchers from the University of Cambridge are among a global team developing a new generation of artificial intelligence models trained not on text or images, but on real-world scientific data, in a move they say could accelerate discovery across disciplines.
The work comes from the Polymathic AI collaboration, which recently unveiled two “foundational” models — Walrus and AION-1 — designed to learn the underlying physics of complex systems. Unlike mainstream AI tools such as ChatGPT, which are trained largely on language or photographs, the new models draw on datasets from astronomy, fluid dynamics and other fields to identify patterns common to multiple types of phenomena.
The Walrus model, led by Polymathic AI researcher Michael McCabe, is trained on the Well, a 15-terabyte dataset compiled from 19 fluid-dynamics scenarios and 63 physical fields. It can analyse systems ranging from exploding stars to atmospheric turbulence and even wireless signal propagation, transferring knowledge from one area to another.
“Training on broader classes makes something easier to use and more likely to generalise,” McCabe said, noting that researchers often lack the time or resources to test multiple models when confronted with unfamiliar physics.
Cambridge mathematician Dr Miles Cranmer said the results were striking. “I continue to be awed by the fact that a multi-disciplinary physics foundation model works at all, let alone at this level,” he said.
The second model, AION-1, is trained on more than 200 million observations from astronomical surveys such as the Sloan Digital Sky Survey and Gaia. By combining images, spectra and other measurements, the system can extract insights from low-resolution data by drawing on patterns learned from millions of other objects.
Researchers liken the approach to human perception, where multiple senses help fill in missing information. The goal is to give scientists a starting point when analysing new experiments, rather than building AI tools from scratch each time.
Dr Payel Mukhopadhyay of Cambridge’s Department of Applied Mathematics and Theoretical Physics said the team hopes scientists will adapt the open-source models for their own work. “Walrus feels like an exciting step towards building a general-purpose AI model for physics-based simulations,” she said.
Polymathic AI principal investigator Shirley Ho said the aim is to embed advanced AI into everyday research. “We want to bring all this AI intelligence to the scientists who need it,” she said.
If successful, the approach could help researchers tackle complex problems faster and with fewer resources, while opening new possibilities for collaboration between scientific fields that rarely intersect.
Looking for your next technology, science or engineering role? Maybe you’re just starting out in your career? Polytec is here to help you find the perfect position for you. Browse our current vacancies now >>> www.polytec.co.uk/jobs
Image credit: University of Cambridge
Researchers at the University of Cambridge have developed a new nanoelectronic device that could significantly reduce the energy demands of artificial intelligence (AI) hardware by mimicking the way the human brain processes information.
The University of Cambridge is set to host the UK’s most powerful quantum computer as part of a major new partnership with US-based quantum technology firm IonQ, marking the University’s largest-ever corporate research collaboration.
The University of Cambridge has been ranked the UK’s top employer in the Financial Times Best Employers 2026 survey, a nationwide assessment based on anonymous employee feedback.
Researchers from the University of Cambridge are among a global team developing a new generation of artificial intelligence models trained not on text or images, but on real-world scientific data, in a move they say could accelerate discovery across disciplines.
A University of Cambridge spinout developing photon-multiplier technology to boost the efficiency of silicon solar panels, has secured a £1.56m pre-Series A investment to accelerate commercial readiness of its breakthrough energy technology.
Developers behind the proposed Cambridge 25 employment park near Junction 25 of the A14 say updated plans could deliver significant economic and community benefits for Greater Cambridge, as a fresh public consultation opens ahead of a spring planning submission.
A vast new employment park could transform Huntingdonshire’s economy, with developers claiming it has the potential to create up to 15,000 jobs.
Cambridge Dictionary has added more than 6,000 new words, phrases and meanings in the past year, with entries ranging from internet slang to political portmanteaux.
The British public is divided over how artificial intelligence should be used in schools, with strong support for its role in helping teachers with administrative work but widespread opposition to its use for marking.
New figures from CV-Library reveal a striking divergence in the UK’s AI job market, with robotics-related positions surging ahead in pay growth while many traditional AI roles experience steep declines.
Oxford and Cambridge universities have each received a £6.25m donation from an anonymous benefactor to expand outreach and support for disadvantaged students in science, technology, engineering and mathematics (STEM).
Meta has launched a new £12m audio research facility in Cambridge as part of its global push to develop advanced AI-powered smart glasses.