AI FOR SCIENCE · INTERVIEWS
Walsh says materials design is moving from 'a million calculations through a funnel' to agents that call tools while they search; his own steps were machine learning in place of solving quantum-mechanical equations, generative models that work backwards from a target property to a crystal, then reinforcement learning, with a warning that the models cheat.
On the record Walsh appears only in one paper's byline, and neither his tie to CuspAI nor the size of its London team is there; the talk adds that he joined this year and that the London office went from two people on 2 January to about 15 at the end of March, and shows which of the methods he presents are his own group's.
No result or number from CuspAI's platform is shown. The generative models and the more than 12 new layered crystals he cites (ten in the matching paper, two of them newly predicted) are his Imperial group's research and a collaborator's synthesis; agents are a direction he names, not something demonstrated.
His starting point: the equations that describe chemistry have long been known but cannot be solved, and a century went into approximations. Machine learning replaces solving them with a learned function; a large quantum-mechanical calculation takes hours to a week, a trained model answers in milliseconds, and he says it is hard to go back. He stresses the prediction part: the aim is a model that does something useful, not only a fit to data. His examples are a solid electrolyte for sodium-ion batteries and the room-temperature behaviour of a perovskite solar cell, both simulations that this kind of model makes feasible.
“once you start to do machine learning it's very difficult to go back to quantum mechanics to be honest.”Aron WalshHe calls the way the input vector encodes chemistry the key choice. Ten years ago features were hand-built (Magpie: one vector per element, each entry with a physical meaning such as atomic number or melting point); now they are learned, for example word vectors trained on a million materials-science abstracts. His group also ran an alignment experiment: using a set of reference materials as anchors, it mapped the representations of models with different architectures and training data into one space, and the agreement was better than he expected. From that he says different models each learn a piece of the puzzle and information can be translated between them.
A crystal is written as atom types, positions and a repeat cell, part categorical and part continuous, which makes modelling awkward. He says generative models made 'work backwards from a target property' feasible for the first time: their first version used contrastive learning on text and a graph neural network as the condition and then generated by denoising diffusion; he says navigating such a composition space would have been a PhD even five years ago and now runs in a day, and deployment-time sampling runs on a laptop. To catch up with Microsoft's MatterGen they moved to a latent space from a variational autoencoder and added reinforcement learning; the latent space alone was worse than real space, and with reinforcement learning it became 'incredibly good', provided the constraints and the assessment are very specific, otherwise the model exploits loopholes and returns small variants of one material, and they spent a lot of effort making the outputs diverse enough.
“the models cheat whenever they can.”Aron WalshHe calls today's usual approach a funnel: put a million materials in, run a great many calculations, and perhaps two candidates come out. He says it works and is not elegant. With this approach, he says a collaborator's group in Korea synthesised more than 12 layered crystals that did not exist before (the matching paper says ten, of which only two came from compounds the screen newly predicted), for instance removing one component from a gallium-and-arsenic ternary crystal to get a layered gallium arsenide unlike the natural form (in the paper the sodium or calcium is etched out, leaving protons or some residual cations). The direction he wants is scientific agents that can call quantum-mechanical calculations, machine-learning stand-ins and generative models and refine from feedback, so that 'we don't have to compute this whole space'.
CuspAI, which he joined this year, is building an agentic platform for materials discovery; it was founded by Chad Edwards and Max Welling, its London office was the two of them on 2 January and is now about 15, with computer scientists, machine-learning experts and chemists and physicists like him. Of the physical-world half he says he did not have time to go deep, and describes only DIGIBAT, which he took part in: a robotic battery platform funded by UK Research and Innovation (UKRI) that makes materials, assembles cells and tests them, fully operating since summer 2025 and now finishing a campaign to produce open battery data, with a comparison of how fast a human expert and the automated system learn.
The record has Walsh only in the byline of the GNoME disorder-classifier paper (2025-10-23), his person page lists only Imperial, and neither the CuspAI entries (the Series B, kUPS, ODAC25) nor CuspAI's page on /funding mention him.
He says he joined CuspAI this year; CuspAI's site and Imperial's release both give his post as chief scientific officer (from January). The talk puts him on the record as a company's science lead and not only an academic.
The kUPS entry (2026-04-23) and the ODAC25 entry cover CuspAI's simulation engine and a generator's output; neither gives team size or how the platform searches.
He says CuspAI is building an agentic platform for materials discovery, and that scientific agents in general can call quantum mechanics, machine-learning stand-ins and generative models; the London team went from two on 2 January to about 15 at the end of March. The specific methods he shows (diffusion, latent space with reinforcement learning) he presents as his Imperial group's, and he does not say which one CuspAI's platform uses.
The only newly made materials he shows are 12 or more new layered crystals, and it came from the funnel he calls inelegant, not from an agent. In the paper only two of the ten are newly predicted compounds, and eight start from known ones. So for an agentic platform such as CuspAI, the evidence he could point to in March came from an academic collaboration and not from the platform itself. He does not blur the two.
Walsh says a model with reinforcement learning finds loopholes and returns variants of one material, and it took a lot of effort to make its outputs diverse. Any claim of 'N new materials generated' can start with: how was diversity measured, and was novelty checked. The GNoME disorder-classifier entry on this record estimates that 80% to 84% of GNoME's predicted compositions are likely to be disordered, the same kind of gap between predicted and makeable.
He says battery processes have many steps and many parameters but very little open data to train on, so DIGIBAT's aim is to produce open data and not first a better model. He adds that materials researchers are paid far less than AI researchers, which makes top people hard to attract, so UK fellowships exist to bring them in. Both are his reading of the UK ecosystem, not CuspAI's conclusions.
A large quantum-mechanical calculation takes hours to a week; a trained machine-learning model infers in milliseconds.
Industry context
A general order-of-magnitude statement, tied to no particular model or benchmark.
Six-component crystals have over 10^13 element combinations (any element allowed), while observed crystals peak at three components.
Industry context
His estimate and his description of the ICSD, with no method given.
Models with different architectures and training data, aligned using anchors, learn atomic representations that agree better than expected.
Current system
His own group's work; the talk gives no metric and no paper.
Navigating such a composition space, a PhD five years ago, now runs in a day.
Guest's judgment
His generalisation; no specific calculation or hardware is shown.
The latent-space model alone was worse than real space and, with reinforcement learning, 'incredibly good'; models exploit loopholes and diversity has to be enforced.
Current system
His own group's result with no numbers and no comparison; he does not say whether this model is used at CuspAI.
A collaborator's group, guided by their screening, synthesised more than 12 layered crystals that did not exist before.
Company-reported
His account of a collaborator's work. The paper is the Nature Materials one of October 2024 (Walsh is a co-author; the author contributions give him and D.W.D. the data-driven models): the screen reached 44 candidates (35 known, 9 unknown) and ten were synthesised, not 'more than 12', and the gap of two is unreconciled. Of the ten, eight start from known ternary compounds and only two (K2In2P3, Na2In2As3) are newly predicted compounds the screen proposed. His gallium arsenide case is, in the paper, Na2Ga2As3 and CaGa2As2 with the sodium or calcium etched out and protons in, layered crystals and not plain GaAs. The full text was read, not the supplementary material.
Agents built for science can call quantum mechanics, machine-learning stand-ins and generative models and refine from feedback without computing the whole space.
Target
A direction; he gives no example of an agent finding a material.
CuspAI's London office was two people on 2 January and is now about 15.
Company-reported
His account of his new employer; no list.
The open-source tool SMACT is downloaded over 5,000 times a month.
Current system
Software his group maintains; the figure is his and the download statistics were not checked.
DIGIBAT has been fully operational since summer 2025 and is finishing a campaign of open battery data.
Current automation
An Imperial facility he took part in; he gives no operating numbers. The dataset was released in July 2026, see the follow-up.
Materials-science researchers are paid much less than AI researchers, which makes top people hard to attract.
Guest's judgment
His observation, with no data.
How far agent search beats funnel screening.
Not yet closed
He gives no comparison and does not say CuspAI's platform has made one.
Whether CuspAI's platform produces a named material it proposed and a partner made and measured.
Would show up as A named material, with the target property, how many candidates were screened, which failed, and the partner lab's measured values.How his 'more than 12' squares with the paper's ten, and whether the rest have a paper of their own.
Would show up as The supplementary material or a follow-up paper listing more than ten crystals and saying how many came from unknown compounds the screen predicted.Partly answered The full text was read (the PDF the author supplied): 44 candidates, eight known plus two newly predicted, ten synthesised in all; the supplementary material was not read and no eleventh or twelfth crystal was found.Whether the latent-space plus reinforcement-learning generator has a public evaluation of the diversity and novelty of its outputs.
Would show up as A paper or benchmark with diversity, novelty and stability measures and a comparison with the real-space diffusion model.Partly answered A November 2025 arXiv preprint by Hyunsoo Park and Walsh at Imperial (2511.07158, main text read, supplementary not read) gives these measures: over 10,000 samples on Alex-MP-20, the latent-space model Chemeleon2 alone is metastable, unique and novel 15.9% of the time against 41.0% for the real-space MatterGen, and 61.3% with reinforcement learning, with novelty at 97.5% and uniqueness down from 99.4% to 88.7%; without the diversity reward its outputs collapse to nearly one composition. But the model and the evaluation come from one group, and stability is scored with a machine-learning force field, not DFT or experiment; no outside group's evaluation of this generator was found.Whether other groups reproduce or compare models trained on the open DIGIBAT data.
Would show up as A paper not written by the DIGIBAT team that uses the dataset and reports the gap between predicted and measured cell performance.First evidence Royce's news release of 13 July 2026 says a dataset of about 250 lithium-ion coin cells is on Zenodo; that is the release, and no paper from an outside group using it was found. The Zenodo record itself was not opened.Later evidence that tests or limits what was said, and targets whose date has passed.
Royce says Imperial's DIGIBAT has released a Discovery Benchmark dataset of about 250 lithium-ion coin cells on Zenodo. This matches the open battery data he promised; it shows the data is out, not that models trained on it work.Henry Royce Institute ↗
He said in March that he had no time to go deep on the physical side; in July the report on CuspAI's Series B says most of the money goes to labs in Cambridge, Singapore and the Bay Area. That answers where the money is going, not whether an agent has found a material.Timeline entry →
A Hong Kong University of Science and Technology preprint gave MatterGen (the model he says their generator was built to catch up with) and forward screening 100 verifications each to find stiff crystals: MatterGen returned 100 hits to screening's 55, but every generated hit contained osmium, 82 repeated a recorded composition, and screening's best candidate was stiffer. It bears out his warning that models cheat and outputs need their own diversity check; on this task the funnel he calls wasteful found the best material. It tests MatterGen, not CuspAI's own model, and it is all computed.Timeline entry →
Read 2026.10.24 This is a conference talk, not a one-to-one interview. What was read is the timestamped full transcript the author supplied (YouTube's machine captions; the audio was not heard, and names are garbled in places, such as 'CASSP' and 'Wuong Shim', so names and titles follow CuspAI's leadership page and Imperial's news release). The video's upload date could not be read: the date is the day of the conference (the organiser's write-up gives 31 March 2026 at the Royal Society), and the length is estimated from the last timestamp in the transcript. What he says in the talk, 'I joined CuspAI this year' and 'on January 2nd there was two of us', matches Imperial's news release of 2026-07-21 (chief scientific officer from January). The organiser lists him under both Imperial and CuspAI, so the page marks which work is his Imperial group's and which is CuspAI's. The layered-crystal paper he alludes to was later found and read in full (Nature Materials, 2024, the PDF the author supplied), without its supplementary material. Open the interview ↗The organiser's write-up of the day (14 April 2026) ↗