How supercomputing and AI are accelerating fusion research

The journal Nature Reviews Physics features a contribution by IPP Director Prof. Frank Jenko on the potential applications of supercomputing and artificial intelligence in fusion research. In the following interview, he explains what these are.

Artificial intelligence—or AI—has long been the tech buzzword of the hour. Thinking machines support us in everyday tasks: generating greeting cards, creating recipes, or helping draft emails. But the field of AI goes far beyond that and has already made its way into the natural sciences—such as cancer research or environmental science. And plasma physics? It too is benefiting from rapid developments: “AI is one of the most promising tools to accelerate fusion research as a whole,” says Prof. Frank Jenko. In his article “Accelerating Fusion Research via Supercomputing”, published in Nature Reviews Physics, and in the interview below, the IPP Director discusses the many ways in which supercomputers and AI are being used in fusion research—and highlights the latest developments, many of which IPP is actively involved in.

IPP: Professor Jenko, how is plasma physics benefiting from advancements in supercomputing and AI research?

Prof. Jenko: There have been rapid developments in supercomputing for many years now, and the equally rapid growth of AI has added another layer. Fusion research benefits immensely from both. Today, thanks to these new capabilities, we can describe real fusion systems with high accuracy—and even extrapolate to systems that haven’t been built yet with a reasonable degree of confidence. This allows us to speed up research and save costs. We believe this makes an important contribution to realizing the dream of a fusion power plant.

IPP: What was the motivation for your article?

Prof. Jenko: The journal’s editorial team approached me. Nature is increasingly focusing on fusion research and identified supercomputing as a particularly important topic. The request came at just the right time, since I feel the same way: Expectations for fusion have increased significantly, both among the public and in politics. If we want to talk about fusion power plants, we also need to talk about how to accelerate their development. Supercomputing and AI can help us here, especially through computer simulations.

IPP: Why simulations in particular?

Prof. Jenko: Because trial-and-error approaches are no longer feasible. In the past, new experiments were often built just to test ideas. But facilities today are much larger and more expensive. Every step—design, preparation, execution, analysis—must be well thought out. Simulations help at every stage. And the models are now so realistic that we can directly compare them to experimental data.

IPP: In your article, you mention that computing power doubles every 18 months. Aren’t we reaching physical limits?

Prof. Jenko: Traditional chips are indeed hitting physical boundaries. But progress continues: GPUs, originally developed for AI applications, are now standard in supercomputing. Our codes have to adapt, but the evolution is far from over.

IPP: How does the development of better computers affect fusion research?

Prof. Jenko: In the past, the goal was often just to describe something qualitatively—that was all that was possible 20 years ago in many areas. In the 1990s, we studied turbulent transport processes using very simple models; even in the 2000s, that was often still the case. But today, we’ve come a long way: We just published another Nature paper comparing detailed ASDEX Upgrade measurements with simulations, and the agreement was very good. That would have been unthinkable twenty years ago. It shows how far we’ve come in being able to simulate reality more and more accurately.

IPP: When did supercomputing first become part of fusion research?

Prof. Jenko: Since the 1960s. The Max Planck Society’s computing center—today’s Max Planck Computing and Data Facility—was founded back then as a department of IPP, specifically for fusion. It was similar in the U.S., where today’s National Energy Research Scientific Computing Center (NERSC) was established in the 1970s. So fusion was an early driver of supercomputing, and still is today.

IPP: But machine learning is a more recent development, right?

Prof. Jenko: The idea is old, but the breakthrough came in the past ten years. Better hardware, more data, and more efficient algorithms have driven the progress. What’s especially exciting for fusion is that we can combine supercomputing and AI. Simulations generate datasets that AI can use for analysis, to accelerate simulations, or for real-time control.

IPP: That sounds like it’s not only relevant for research, but also for operating future power plants?

Prof. Jenko: Absolutely. Especially for real-time control, such as avoiding disruptions, AI is indispensable. This has to happen automatically and in milliseconds—humans can’t do that.

IPP: Are existing codes like GENE being adapted?

Prof. Jenko: Yes. Switching to GPUs is a major effort. In the past, switching to a new computing architecture took days or weeks; today, it can take months or even years. But it’s worth it: Like in Formula 1, if you want to win the race, you need the best engine.

IPP: Does AI always need supercomputers, or can it run on regular computers?

Prof. Jenko: It depends. Training data has to come from somewhere, and that often means simulations—which require a lot of computing power. But once a model is trained, it can run on smaller machines. That’s actually necessary for real-time applications.

IPP: What role does MPCDF play today?

Prof. Jenko: It remains a key partner in supercomputing and increasingly in AI as well. The newest AI algorithms are developed elsewhere, but MPCDF helps make them usable for fusion research.

IPP: Where are the most important AI development centers?

Prof. Jenko: Universities like TU Munich, and also international tech companies. The U.S. and China are making major investments—sometimes driven by commercial interests. At IPP, we focus on fusion-relevant applications and on providing the data and questions.

IPP: And what about licenses for those algorithms?

Prof. Jenko: Many are open source—for example, Python libraries. But there’s also commercial software. We collaborate with Google DeepMind, for example, and licensing can be a bit more complex there. Whenever possible, we use freely available tools.

IPP: What role do supercomputing and AI play in predicting plasma performance?

Prof. Jenko: A central one. High-accuracy simulations require a lot of computing time. But optimization processes need faster models. So we’re building databases and training neural networks on them. I call this a “multi-fidelity approach”: accurate models working hand in hand with simplified, faster ones. Both are important.

IPP: Can machine learning help prevent disruptions?

Prof. Jenko: Disruptions and the runaway electrons they produce can damage a tokamak. AI can act as an early warning system and automatically initiate countermeasures. It’s a big challenge, especially because of the required reliability.

IPP: Are supercomputing and AI also useful for stellarators?

Prof. Jenko: Yes, though the focus is different. Stellarators offer many degrees of freedom in magnetic field design. The design of Wendelstein 7-X would have been impossible without supercomputing. AI can help optimize these designs more efficiently—for example, to better control turbulent transport.

IPP: Is AI also helping in fusion-oriented materials research—and if so, how?

Prof. Jenko: Materials are affected by physical processes across many spatial and temporal scales, from macroscopic structural changes to atomic-level effects. Only complex models can capture all that. AI can help pre-select promising material candidates. Experiments are still needed, but they can be carried out more efficiently.

IPP: What are digital twins, which you also mention in your article?

Prof. Jenko: A digital twin is a computer model of a real system. It allows rapid and cost-effective testing and optimization. In fusion, digital twins range from individual plasma processes to entire power plant systems. It’s better to think of them as a family of digital twins.

IPP: And IPP is leading in this area?

Prof. Jenko: Absolutely! Many of our tools are being used in the development of digital twins. Overall, I’d say that especially in plasma simulation, there has been tremendous progress in recent decades. Today we can simulate with a level of accuracy that would have been unimaginable in the past.

Prof. Jenko’s article “Accelerating Fusion Research via Supercomputing” can be found at this link: https://www.nature.com/articles/s42254-025-00837-1