Lorenzo Zanisi, Lead Data Scientist at the UK Atomic Energy Authority, says that to design the STEP fusion power plant his colleagues use AI to join messy experimental data to high-fidelity simulation: a multimodal foundation model, an open-source plasma shape simulator and a set of surrogate models that cut one trajectory simulation from four weeks to four hours and one simplified MHD simulation from 354 hours to about 10 milliseconds.
STEP will run plasmas no experiment has seen, so he says it needs top-fidelity simulation and an AI digital twin
UKAEA's STEP, the Spherical Tokamak for Energy Production at West Burton, is meant to be the first fusion power plant and to put power on the grid in the 2040s, Zanisi says, and STEP and adjacent programmes have just been awarded about £2.5 billion to 2030. A power plant will run plasmas no experiment has reached, so they must be simulated at the highest fidelity on the most powerful supercomputers. AI's opening, he says, is to combine messy experimental data with high-fidelity simulation into a predictive model for design, control and real-time forecasting, under the name digital twin; Sunrise, an AI supercomputer just for fusion, is due towards the end of 2026, he says.
They fold every diagnostic into one multimodal foundation model, and pretraining helps most on the hardest, high-frequency tasks
A tokamak plasma is mostly too hot to see, so it is measured by many diagnostics hidden in the wall; in a power plant, he says, they will produce about a petabyte a day. The data comes in several modalities sampled at different rates, from one-dimensional time series to video. Rather than a model for each, the team built, with partners including IBM, a prototype multimodal foundation model: a multimodal encoder and a transformer backbone. Fine-tuning from a pretrained representation beats training from scratch, he says, and the gap is largest on the hardest tasks, those full of high-frequency information.
Fusion data was never collected for AI, so they first built the FAIRMAST dataset
That foundation model was slow in coming, he says, because fusion data was never collected with AI in mind: it is not catalogued for AI to search, a stretch of experimental data sometimes cannot be identified, signals are mislabelled, and the data sits in systems inside systems. Climate has ERA5; fusion had nothing like it. So they built FAIRMAST, which he calls the first fusion dataset to follow the FAIR principles (findable, accessible, interoperable, reusable), led by Samuel Jackson. It is not fun work, he says, but all the AI rests on it, and more such fusion datasets are coming.
Inspired by DeepMind's reinforcement-learning plasma control, they built the open-source FreeGSNKE simulator and are putting it into the control system
Plasma shape is hard to control, he says, and decides whether a power plant gets energy out. Four years ago DeepMind trained a reinforcement-learning agent offline on simulators and used it to control plasmas in Switzerland's TCV tokamak, reaching shape trajectories control engineers find very hard. UKAEA then started its own AI control project, building FreeGSNKE, an open-source simulator of how plasma shapes evolve, and accelerating its key components so experiments can be designed much faster without breaching the magnets' current limits. These were going into the control system that month, March 2026, he says, and he could not say more.
Spend a month or two computing a big simulation database, train a surrogate, and one trajectory simulation drops from four weeks to four hours
Traditionally, he says, physicists run a few simulations at a time with slow, brittle codes, and every new point in parameter space means solving again, sometimes waiting a week. The new workflow accepts the cost: take several cluster nodes for a month or two, generate as many simulations as possible, train a surrogate on that database, and answer each query in milliseconds. Turbulence makes the core colder than wanted, and one profile prediction takes about a thousand high-fidelity simulations, too many; empirical scaling laws do not extrapolate to a power plant. A simple multilayer perceptron surrogate already reproduces measured temperature profiles fairly well, and inside a trajectory simulation it cuts one run from four weeks to four hours. He also shows a surrogate that failed, warning that they fail if built carelessly.
A surrogate for five-dimensional turbulence runs about a million times faster than simulation and about four times more precisely than the low-fidelity model
High-fidelity turbulence is described by the gyrokinetic equation, which evolves a distribution function in five dimensions and is very hard to solve. With Johannes Kepler University in Linz, he says, UKAEA built the first surrogate that can process that five-dimensional data: about a million times faster than simulation once deployed and about four times more precise than the low-fidelity model, though not as precise as the full simulation, so a new level in between. Keeping all five dimensions lets it give turbulence spectra where there was no data and reproduce zonal flows. To cut costs, active learning saved up to a factor of 20 in data generation on simpler simulators, and neural-field compression shrinks a two-terabyte simulation up to 70,000 times while keeping the physics, in one of the best examples, he says.
Edge simulators are old and brittle, and a new simulation manager has run about 40,000 of them, twice JET's twenty-year total
The core must reach a hundred million degrees, yet the plasma cannot touch the wall at that, since materials melt at about five thousand, and core to edge is about a metre. The edge is a huge multi-physics problem whose equations carry parameters nobody knows, picked out of thin air or calibrated on experiments, which a power plant design cannot rely on. The simulators are old and break all the time, so the team built a simulation manager with machine learning and a human in the loop that has run about 40,000 simulations to date; JET's whole twenty-year history saw about 20,000, he says. A simple transformer surrogate predicts wall temperature almost indistinguishably from the truth, showing at a glance where operation stays under the 5,000-degree line, the Pareto fronts of design trade-offs, and uncertainty over the unknown parameters, which he says was never done before.
One MHD simulation takes 354 hours and a surrogate rollout about 10 milliseconds; a PDE foundation model matches it on a tenth of the data
A hundred million degrees in the core and five thousand at the edge within a metre makes for extreme gradients and magnetohydrodynamic instabilities, which must at least be known about and ideally controlled. The team built a surrogate of a simplified MHD system that is close to the ground truth as the instabilities start and statistically accurate in the messy phase after; one simulation took 354 hours, one full rollout about 10 milliseconds on a GPU. They also tried, zero-shot, a PDE foundation model released about three years ago that had never seen such instabilities, and it already did well; fine-tuned on 10% of the data they used before, it matched the same performance. That is an argument for data efficiency, he says, since their simulations are too expensive to make many.
The figures are UKAEA's own, many on simplified systems, and the power plant date is a target
This is Zanisi's talk at The Alan Turing Institute's AI for Science event, on work by UKAEA and its partners, with no outside evaluation, and he says he was not personally involved in some of it. The active-learning savings come from simpler simulators, the 70,000-fold compression is one of the best examples, and the MHD surrogate covers a simplified system; what goes into the control system and how it performs he cannot yet say. STEP on the grid in the 2040s is a target, and he says himself that fusion is always twenty years away. The talk was given on 31 March 2026, and its figures are as of that day.

