UK and US fusion supercomputers team up to build shared AI models

Two fusion supercomputers in the UK and US could be linked to train the same artificial intelligence models on experimental data from different fusion devices, potentially helping engineers design future power plants more quickly. The proposed SUNRISE-STELLAR-AI alliance would connect the UK Atomic Energy Authority's (UKAEA) SUNRISE supercomputer with the STELLAR-AI platform at the Princeton Plasma Physics Laboratory (PPPL). Researchers will use data from the MAST Upgrade facility in Oxfordshire and the NSTX-U facility at PPPL in New Jersey.

Both facilities are compact spherical tokamaks with similar designs, a similarity that helps researchers combine experimental results to train models better able to predict fusion plasma behavior across different machines.

The partnership was announced on September 14 at a global fusion policy summit in London and builds on a memorandum of understanding signed between UKAEA and PPPL in June. The parties are still in the exploratory phase of how to connect the computing platforms.

Goals and significance of the alliance

The proposed alliance would allow the two supercomputing platforms to train identical models using data from both national laboratories. This matters because models trained only on results from one fusion machine can struggle when applied to another. By combining datasets, researchers hope to build machine learning systems that capture more of the underlying physics and make more reliable predictions. The computers could also fill gaps in experimental datasets through simulation, expanding what researchers can study without running every scenario on physical machines.

Joe Milnes, executive director for engineering and computing at the UK Atomic Energy Authority, said: "Fusion is one of the great scientific and engineering challenges of our time. To tackle those challenges, fusion needs partnerships. The US and UK are global leaders in fusion research, and a SUNRISE and STELLAR-AI alliance will build on a long history of transatlantic collaboration to advance fusion further."

SUNRISE is the UK's first AI-focused supercomputer for fusion energy, backed by £45 million in government funding. STELLAR-AI provides PPPL's AI and high-performance computing capabilities. The planned system also aims to support the development of digital twins for both tokamak facilities.

Digital twin applications

Digital twins would create detailed virtual models of fusion machines based on experimental data. Researchers could use the models to test changes and predict machine behavior in software before applying them to physical equipment. Rob Akers, director of computing programs at UKAEA, noted that the two laboratories could combine experimental data with simulation to explore untested operating conditions. "We can jointly develop digital twins of both machines to support the design of future fusion power plants, creating more predictive and workable spherical tokamak models, ultimately more practically useful for fusion engineers," he said.

The alliance could ultimately support work on future projects, including the UK's Spherical Tokamak for Energy Production (STEP Fusion) and the proposed advanced spherical tokamak reactor at PPPL. Researchers also plan to study how to move computing tasks between the two platforms despite their different hardware. This would allow scientists to pick the system best suited to a particular computational task. Shantenu Jha, head of computational science at PPPL, said: "Our goal is to enable models and experiments to flow freely between the two systems.

We will turn a set of supercomputers into a single engine for fusion discovery."

The proposed alliance has not yet become an operational shared supercomputer, with partners still working through its technical requirements and next steps.

UK-US collaboration shapes the future of fusion research

UK-US collaboration in fusion research represents a joint effort by the global scientific community to address energy challenges. By pooling supercomputing resources from both countries, researchers can train AI models more effectively, which is essential for designing future power plants. Digital twin technology will enable researchers to conduct experiments in virtual environments, reducing the risks and costs of physical operations. This cross-border collaboration could not only accelerate technological progress but also open new opportunities for the global energy transition.