Could better mathematics help AI become a more useful partner in scientific discovery?
That is the ambition behind a £2 million research fellowship awarded to Lancaster University mathematician Dr Henry Moss.
The university’s Data Science and AI Institute newsletter, published on 2 November, confirms that Dr Moss has received a UKRI Future Leaders Fellowship to develop the mathematical foundations of generative AI for scientific discovery.
It is positive news about backing British research. But the award is funding for work to be done, not proof that a new treatment, scientific breakthrough or commercial product has already been delivered.
What the £2 Million Fellowship Supports
Dr Moss is part of Lancaster’s Mathematics for AI in Real-world Systems group, known as MARS, in the School of Mathematical Sciences. The university describes his work as fundamental AI research alongside developing machine-learning tools and software for real-world applications.
The fellowship focuses on the mathematical underpinnings of generative AI. These are the principles that help researchers understand how such systems work and how they might be made useful for scientific problems.
Generative AI is often associated with creating text or images. In science, related methods can also be investigated as ways to propose candidate structures, reconstruct information or explore possible solutions to a difficult problem.
The important word is “investigated”. Generating a plausible-looking answer is not the same as establishing that it is physically valid, experimentally useful or safe. Mathematical foundations matter because scientific tools need more than convincing outputs.
Why AI for Science Needs Careful Foundations
Imagine a researcher exploring possible molecular structures. A system that proposes many candidates may be useful, but only if there are reliable ways to assess the proposals and decide which deserve further work.
Or consider an imaging problem where a computer tries to reconstruct missing information. A visually convincing reconstruction is not automatically an accurate account of the original object.
These are examples of the kinds of questions that make this research direction important, not specific achievements claimed for Dr Moss’s fellowship. The newsletter confirms the award and its scientific aim; it does not establish that those possible applications have been solved.
Work on uncertainty, physical constraints and the behaviour of learning systems can help researchers ask better questions about when an AI result deserves confidence and when it needs further checking.
Ten Lancaster Papers Accepted at NeurIPS 2026
The same newsletter reports a second research milestone: ten Lancaster papers accepted at the main conference of NeurIPS 2026, a major international venue for machine-learning research.
Seven come from MARS. The remaining three are from the Multimodal Perception and Intelligence Learning laboratory, or MPIL, in the School of Computing and Communications.
The newsletter also mentions six further MARS workshop papers. Those are separate from the ten main-conference acceptances, so they should not be mixed together when describing the headline total.
These acceptances describe research across the groups. They are not ten completed products, and they should not all be presented as outputs of Dr Moss’s newly announced fellowship.
From Molecular Modelling to Radiology and Robotics
The range of topics shows why AI research cannot be reduced to chatbots.
- AI for science: MARS work includes ways of conditioning Gaussian processes on physical laws and language, alongside diffusion methods for image reconstruction.
- Molecular modelling and search: active learning and Bayesian optimisation feature in work on selecting informative examples and exploring difficult design problems.
- Medical imaging and radiology reports: the programme includes image-reconstruction research and MPIL work on clinical reasoning for radiology report generation. These are research topics, not evidence of improved care.
- Robotics: an MPIL paper examines task-oriented dexterous grasp generation, including constraints intended to address physical plausibility.
- Reliability: a MARS paper investigates conformal prediction when calibration data are contaminated, a technical question relevant to the dependability of uncertainty estimates.
These examples come from the university’s descriptions of the accepted work. They show the breadth of the research, rather than proving that every method will transfer successfully into everyday use.
What This News Does and Does Not Demonstrate
For any medical application, further validation would be needed to establish whether a method works reliably with different patients, equipment and clinical settings. Good benchmark results or an accepted paper cannot substitute for that evidence.
For robotics and scientific design, there is also a gap between a research method and a dependable tool. Testing, engineering and real-world evaluation remain essential.
The good news here is that British researchers are receiving support to tackle those underlying challenges, while contributing work across science, imaging, robotics and reliability.
Not every important AI story begins with a finished system. Sometimes it begins with giving researchers the resources to find out what works, what does not, and what would make the technology more trustworthy.
Sources and Further Reading
- Lancaster University: DSAIL Newsletter, November 2026, published 2 November 2026. Source for the £2 million fellowship, Dr Moss’s research aim, and the ten main-conference paper acceptances and their topics.
- MARS: NeurIPS 2026 papers, linked by the university newsletter. Read alongside its distinction between seven main-conference papers and six workshop papers.
- MPIL laboratory publications, linked by the newsletter for the robotics, radiology-report and visual-generation research.
Less hype. More real-world AI progress.