AI can predict, but only experiments can prove: combining computation, mathematics and neutron spectroscopy
14 Jul 2026
As neutron spectroscopy instruments become more advanced, and the systems studied more complex, the data analysis tools required must also evolve. AI and machine learning offer powerful new ways to model materials and extract richer information from experimental data, but in the physical sciences their predictions are only as trustworthy as the experiments used to test them.
ISIS instrument scientist Jeff Armstrong has been developing methods that combine computation, mathematics and neutron spectroscopy, demonstrating that AI and experimental science are not competing routes to discovery, but essential partners.
AI models for understanding chemical processes
Jeff has been working with colleagues from STFC’s Scientific Computing department to use neutrons to carry out one of the most demanding experimental tests of whether AI can reliably predict atomistic dynamics. Their paper, published in The Journal of Physical Chemistry Letters, is about machine-learned interatomic potentials; AI models that learn how atoms and molecules interact.
Once trained, these models can run incredibly quickly, sometimes on a laptop, unlike traditional quantum-mechanical simulations that often need large supercomputers. This could be hugely important for understanding processes in batteries, chemical reactions and porous materials, where subtle atomic-scale motion controls how well materials work.
The challenge is that these AI models are usually tested against other calculations, rather than directly against experiment. In this work, Jeff used a new high-pressure cell on Tosca to squeeze a molecular crystal and measure how its vibrations changed. This provided data he could use with the team from Scientific Computing to test two generations of machine-learned interatomic potential models, run using STFC’s scarf high performance computing facility.
They found that this method worked well for validating the models, which were able to predict the vibrational behaviour of the molecular crystal both with and without applied pressure.
“This gave us a particularly tough experimental test of whether the AI model was really predicting the forces between atoms correctly,” explains Jeff. “This shows that neutron spectroscopy can help make AI simulations of materials trustworthy.”
Mathematical tracking of the effects of isotopic substitution
In another recent publication, Jeff developed a mathematical method of tracking how vibrational modes change when hydrogen is replaced with deuterium. He used data from an inelastic neutron scattering (INS) investigation on Tosca into the protiated and deuterated forms of an industrially relevant MOF, ZIF-8, and combined them with density functional theory lattice dynamics (DFT-LD) calculations.
His new method modelled the vibrations of the molecule as they change across the full theoretical space between the masses of hydrogen and deuterium. This revealed a surprisingly complex interaction, with the peaks not just shifting, but changing in distribution and intensity.
Although commonly used in other techniques to selectively ‘hide’ parts of a system, selective deuteration is not commonly used in INS. This paper shows that this is for good reason: as deuteration actually changes the energy of the vibrations themselves, sometimes in unexpected ways.
However, this result brings exciting potential applications in organic conducting materials, which are used in wearable devices. Because molecular vibrations are key to how electrons are transported, it could be possible to selectively deuterate these compounds to engineer them to have more desirable electronic properties.
Using simulations to analyse QENS data more reliably
In collaboration with the universities of Bristol and Oxford, Jeff has also been working on another new analysis method to extract molecular motion from quasi-elastic neutron scattering (QENS) data more reliably.
Typically, QENS spectra are analysed by fitting a selected number of Lorentzian functions independently, often relying on oversimplified models. This means that they often fail to determine the underlying dynamics correctly. As catalytic and functional materials and processes grow in complexity, rigorous discrimination between competing transport mechanisms becomes even more important.
In this study, the group have developed a new integrated model for analysing QENS data. Recent studies at ISIS have found that the coherent contribution to QENS experiments should not be ignored, and so as well as using combined incoherent/coherent scattering data collected on the Iris instrument, the group used polarised neutrons on LET to decompose the incoherent and coherent scattering data in order to give context to the model.
They used their model to demonstrate that QENS can, for the first time, resolve the different types of rotation occurring in liquid benzene, a prototypical aromatic molecule relevant to microporous catalysis.
“Although a classic, well-studied system, we were able to use our model to extract previously unobtained information about the dynamics,” explains Andrew McCluskey from the University of Bristol.
The ability to discriminate between different types of rotation has implications for catalysis, as these rotations are often critical to the rate-limiting steps of reactions. The next step for the team is to develop their model into software that other neutron scatterers would be able to apply in their work.
Integrating computing into spectroscopy
These three recent studies are just a few examples of how modern computation, mathematics and neutron spectroscopy can be combined to understand atomic and molecular dynamics in much greater detail. They show the development from traditional peak fitting towards more rigorous, model-aware ways of gaining insight from neutron data.
“In addition, many of these approaches could, in principle, become more automated, helping users extract deeper physical information from neutron experiments in a more systematic and reproducible way,” explains Jeff.
