In the quest to replace fossil fuels with cleaner forms of energy, few options are as alluring as nuclear fusion, the reaction powering the Sun and all the stars in the universe. Fusion is carbon-free, and unlike solar or wind, it can be harnessed continuously, day or night.

To make fusion work on Earth, hydrogen isotopes are fused at temperatures hotter than the Sun’s core to form plasma in a kind of magnetized bottle called a tokamak. Extracting useful energy from this ring of superheated plasma is a highly challenging control problem: the plasma must be kept hot and dense enough to release more energy from the reaction than went in — a holy grail known as ‘breakeven,’ while maintaining enough stability to avoid damaging the machine.

More than 50 tokamaks are in operation worldwide, with the eventual goal of kickstarting commercial fusion and delivering net fusion energy to the grid. Among the key players is the UK Atomic Energy Authority (UKAEA), which has collaborated with IBM Research and STFC Hartree Centre to use the latest algorithms to transform raw tokamak data into a form that can unlock new insights into plasma behavior.

Together, IBM and its collaborators just open-sourced TokaMind, a first-of-its-kind AI model for fusion plasma that harmonizes sensor data of varying frequencies and timescales from the UKAEA’s Mega Ampere Spherical Tokamak (MAST) into a structured, unified representation of how plasma responds to a tokamak's operational settings.

“We set out on a journey last year with the UKAEA and STFC to explore how the most advanced foundational AI techniques could support the modeling of fusion plasma and its applications,” said Juan Bernabé-Moreno, Director of IBM Research Europe for Ireland and the UK. “Fast forward two years, we have not only released the first AI foundation model for a tokamak, but set new standards in algorithmic research for fusion.”

TokaMind currently represents just one tokamak, but UKAEA scientists plan to contribute data from other machines, allowing them to compare results and pick out the most promising designs. They also plan to integrate mathematical descriptions of plasma behavior, validated through simulation, into the model to find new ways to control and amplify fusion performance.

“We could potentially augment existing experiments with new actuators to poke the plasma and make it perform better – to take it into a new regime of operation,” said Rob Akers, director of computing programs at UKAEA. “If we start combining experimental data with our theoretical or ‘model’ based understanding of the plasma, it could be transformational.”

In an area south of Oxford, England, bordered by farms and the river Thames, scientists at the UKAEA’s MAST Upgrade (MAST-U) are busy running experiments, or “shots,” to learn how they can better tame this promising energy source. Just a few seconds long, each shot records how the confined plasma responds to different manipulations. Experiments underway now will influence the design of the UKAEA’s prototype fusion power plant, Spherical Tokamak for Energy Production (STEP), when it comes online sometime in the 2040s.

“The big question is whether we can confidently extend these models to the point where they can help us design commercial-era power plants,” said Akers.

Modeling fusion plasma behavior

Researchers trained TokaMind on data from the MAST experiments, which ran from 1999 to 2013 and confirmed that a compact machine, shaped more like a cored apple than a donut, could be a viable fusion power plant design. TokaMind is expected to produce new insights that will help engineers refine STEP’s design and guide its initial operation.

Until transformer-based AI models came along, scientists had no real way to unpack the full variety and scale of data streaming from MAST’s sensors. Fusion science breakthroughs historically have come from a subset of data that scientists found most compelling.

“The temptation is always to look at the most interesting, highest-performing plasmas,” said Akers. “I would argue that there might be equally important information in the shots that look a bit more boring.

“Now that we can build models around the entire corpus of data, I suspect we will learn valuable new information,” added Akers. “AI will allow us to make decisions based upon all the data.”

Like previous AI models IBM built with NASA and the European Space Agency, condensing and harmonizing vast amounts of geospatial, solar, and climate data, TokaMind has proved capable of transforming raw MAST data into knowledge that can be transferred to a variety of plasma-analysis tasks.

TokaMind has been put through its paces on these tasks in a new benchmark, also designed and open-sourced by the researchers, called TokaMark. On nearly all 14 tasks, the foundation model did better than a traditional machine-learning model trained for each task independently.

Pre-training itself seems to have made a difference. Like previous models, TokaMind was given partially masked data that it had to reconstruct during training. In the process of learning to fill in the blanks, the model seems to have gained knowledge that helped it most on the hardest tasks.

TokaMind had the clearest edge over the baseline model on long-term forecasting tasks. “It really seems to have helped the model learn a transferrable representation of plasma dynamics,” said Tobia Boschi, an IBM researcher who helped train the model. “We wanted to solve all our tasks with one model and a light fine-tuning — and we did.”

Masked pre-training also helped improve the quality of the MAST dataset generally. In a typical tokamak experiment, hundreds of sensors gather data from each shot, and sometimes glitches occur. “You don’t want to get rid of the shot because one sensor wasn’t working,” said Alessandra Pascale, an IBM researcher who led the engineering team. “The model effectively helped us to salvage the data.”

At 9 million parameters, TokaMind is small by generative model standards, but the complexity of its training data may be unrivalled. About 40 different signals of varying frequencies are fused and encoded in the model, from physical descriptions of the plasma to diagnostic data from the machine and its magnets.

In the next phase of the project, researchers will expand TokaMind to include data from MAST-U and potentially other tokamaks. With more data, the model could provide insights into how the machine's design and other key variables influence plasma state, creating a new lever for optimizing machine geometry and operations.

Harnessing the power of stars on Earth

The Apollo mission’s true benefit came not from putting humans on the moon, Akers likes to argue, but in the spin-off technologies that came after. “Low-Earth orbit satellites, microelectronics — it was all dragged forward by the Apollo mission,” he said. “It didn’t matter whether Neil Armstrong put his footprints on the moon, what he did was give human beings the audacity to do big things. And fusion is a big thing.”

To harness the power of stars here on Earth, scientists will need to combine their most powerful tools. In fusion science, that includes cutting edge algorithms to bring vast experimental datasets down to size, and mathematical models, based on partial differential equations (PDEs), that can simulate the complex interactions between a tokamak and its plasma. The team envisions that powerful classical and quantum computers could also be needed to run these models and safely take the final leap safely to commercial fusion.

Foundation models now give scientists the chance to combine insights gained from both real-world experimental data and physics-based simulation data. In the next phase of the IBM-UKAEA collaboration, researchers will introduce simulation data into TokaMind, taking advantage of both perspectives. The UKAEA’s new AI supercomputer, Sunrise, soon to be the most powerful supercomputer dedicated to fusion energy, will play a key role as the project continues.

“We can now build surrogate AI models that represent these complex systems,” said Akers. “They may not be as accurate as a full-blown, high-fidelity simulation, but they can allow us to explore more design-space options and identify uncertainties in our predictions.

“By combining AI surrogate models with high-fidelity simulations, we can improve the overall quality of our predictions and avoid highly unlikely but potentially catastrophic ‘black swan’ events.”

One of the first equations that researchers have targeted for TokaMind is the Grad-Shafranov force-balance equation for equilibrium in a tokamak plasma, where the plasma’s outward pressure is balanced by the inward force of the magnetic field confining it. The equation is key to understanding plasma shape and position, that with other formal knowledge, could give the model a better understanding of plasma state and improve its predictive capabilities.

What types of plasma ultimately produce energy most efficiently is still an open question. TokaMind is a first step toward evaluating the different ideas being explored. “There’s still a lot of heavy lifting to do before we have a model we can just roll out onto the experiment and start using,” said Akers. “But it’s been a great year, and our collaboration with IBM and Hartree Centre is helping us stay focused on achieving fusion’s moon-landing moment.”

Cracking fusion plasma modeling

AI surrogates are faster and cheaper than running full physics simulations at scale, but other technologies will also be needed to crack plasma modeling for fusion.

"TokaMind will deliver tremendous value to the field, but there are still complex scenarios where statistical learning alone falls short,” said Bernabé-Moreno. “We will need a new generation of algorithms — ones that can exploit fundamentally different ways of representing information.”

Plasma behavior is governed by atomic-level interactions, turbulence at many scales, and highly nonlinear dynamics. Data-driven models like TokaMind can pick out patterns in simulation data, but they inherit the limitations of that data, too. When the underlying physics is too complex to simulate using classical computation methods, accuracy eventually plateaus.

Quantum computing, by contrast, can natively represent quantum-mechanical systems — things like atoms, electrons, and photons — using wave functions within an abstract Hilbert space describing all potential states. For plasma, quantum algorithms open the door to simulating kinetic and many-body effects more faithfully at their natural scale. This can generate higher-fidelity training data that AI surrogates can use to tackle parts of the problem by solving partial differential equations, or sampling complex distributions.

“The computational challenges we face in fusion can only be solved by merging quantum computing, AI, and high-performance computing (HPC),” said Alessandro Curioni, an IBM Fellow and VP of Algorithms and Applications at IBM Research.

“Quantum can handle the physics that classical computing cannot while AI can deliver fast and reliable results; HPC is the computational backbone that can scale and integrate these complementary methods,” added Curioni. “Together, they can push the frontier of plasma modeling.”