Researchers at the U.S. Department of Energy’s Princeton Plasma Physics Laboratory have cleared a major hurdle in their pursuit of commercial fusion energy. In a paper published in nature communications In May 2024, lead author SangKyeun Kim and colleagues reported that machine learning software successfully prevented plasma instability in two separate tokamaks (DIII-D and KSTAR) operating in high confinement mode. This is the first time this has been achieved under conditions directly relevant to commercial-scale fusion power generation.
Machine learning controls plasma instability across two tokamaks
May 2024 nature communications This paper details how PPPL researchers deployed the same machine learning code on two different tokamaks (a DIII-D device in the United States and a KSTAR in South Korea) and achieved stable high confinement modes in both without introducing instability.
“This result is particularly impressive because we were able to achieve the result in two different tokamaks using the same code,” said lead author SangKyeun Kim. This point is really important. Fusion devices vary in design, scale, and operating conditions; Single transportable codebase This represents meaningful progress toward general-purpose plasma control, which the field has long needed.
This is the first time researchers have achieved this result under conditions directly relevant to commercial fusion power generation. There is a large gap between laboratory demonstrations and commercially relevant performance, and closing this gap moves the technology closer to practical deployment.
Why plasma instability is a major challenge for nuclear fusion
Fusion plasma has almost no errors. Conditions inside a tokamak change every millisecond, so continuous management is required. A human operator cannot respond at that speed.
Failure to control can result in disruptions of sudden magnetic perturbations that destabilize the entire plasma. These are not just inefficiencies. Destruction can cause severe physical damage to the fusion vessel, which is unacceptable in commercial reactors designed to operate reliably for many years.
This challenge is most severe in H mode. Highly confined mode is the plasma state necessary for commercial power generation, but it is also the most difficult state to stabilize. At the edge of the plasma, an explosive instability occurs called . Edge localized modeor ELMs, can explode without warning, drawing impurities from the vessel walls into the plasma and reducing fusion efficiency. Suppressing the ELM while preserving the H mode is one of the central problems in fusion engineering.
How machine learning systems work
Traditional control code follows fixed instructions. Machine learning works differently. Analyze your data, identify relationships between variables, and adjust your responses based on what you learn. This adaptability lends itself to real-time plasma control.
The system developed at PPPL (built by associate professor and PPPL researcher Egemen Koremen) can predict when disruptions will occur, determine which parameters need to be adjusted, and make those changes before instability occurs, all within milliseconds. Don’t wait for problems to appear. Predict and correct for events in advance.
PPPL has been building toward this capability for years. Back in 2019, principal research physicist William Tan and his team demonstrated for the first time the transfer of a fracture control model from one tokamak to another. The work published in natureestablished the conceptual foundation for what Kim’s team is currently achieving at a commercially relevant scale.
Extensive AI applications in fusion research at PPPL
The research on plasma stability is part of PPPL’s broader effort to apply machine learning across multiple areas of fusion research, each targeting different computational bottlenecks.
Michael Churchill, head of digital engineering at PPPL, uses machine learning to accelerate optimization of stellarator designs. Stellarators are more geometrically complex than tokamaks and require multiple simulation code runs during design validation. Some of them, such as the widely used XGC, require advanced supercomputers, but still don’t run very fast. Machine learning can help close that gap by enabling faster, higher-fidelity computation.
Another team is applying AI to the HEAT code, which models heat flux within tokamak diverters. Researcher Domenica Corona Rivera has already significantly reduced the calculation time while keeping the results loose. 90% consistency In the original code. The goal is to run the code between plasma shots and use the results to adjust the parameters for the next discharge.
Associate research physicist Alvaro Sánchez Villar and his team are using machine learning to optimize ion cyclotron radiofrequency heating. Their accelerated model produces results in microseconds instead of minutes, with minimal loss of accuracy. This difference makes real-time control applications less theoretical and more feasible.
Rounding out the effort, Principal Research Physicist Fatima Ebrahimi will lead a four-year DOE-funded project that combines experimental data, validated simulations, and machine learning to study plasma edge behavior. The goal is to identify strategies for confining plasma in commercial-scale tokamaks that are larger and operate at higher temperatures than today’s experimental equipment.
Important points
The PPPL findings are a synthesis of several firsts. The same machine learning code worked on two different tokamaks, both operating in H mode (the commercially required plasma state), avoiding the instabilities that have historically made that state very difficult to maintain.
PPPL’s extensive AI program suggests that this is not an isolated result. Machine learning is being applied to stellarator design, heat flux modeling, plasma heating, and edge confinement, with each effort targeting specific obstacles on the path to commercial fusion. 2024 nature communications This paper is one data point in a larger, ongoing effort to make fusion reactors controllable, durable, and ultimately viable at scale.
Carlos is an engineer with strong expertise in technical and industrial topics. He previously worked for international companies such as Siemens and speaks Spanish, German, English and Italian.
