AI Controls Fusion Plasma Faster Than Humans Can React

The framework, called PACMAN (Prediction And Control using MAchiNe learning), has been successfully tested in five separate experiments on a real fusion device. Its design and early results are described in a new paper published in the journal Nuclear Fusion.

AI addresses fusion plasma control challenges in milliseconds

Nuclear fusion could provide a virtually unlimited source of low-carbon energy. Researchers are investigating several approaches to making fusion practical on Earth, including tokamaks. These machines use powerful magnetic fields to confine plasma, an electrically charged gas often described as the fourth state of matter.

For fusion reactions to continue, plasma must remain extremely hot, dense and stable. Maintaining those conditions requires constant adjustments to tokamak heating systems, magnets, gas injectors and other equipment. Even small disturbances, known as plasma instabilities, can grow within milliseconds and interrupt the fusion reaction.

Predicting plasma behavior is another major challenge. Advanced computer simulations can require days or even months to complete. While these simulations are valuable for planning future experiments, they are too slow for real-time control when an experiment lasts only a few minutes.

“This is great for preparing for the next experiment in a year, but for control we need a model that makes instantaneous decisions,” said co-lead author Hiro Fareh Kaga, a graduate student in the Princeton Program in Plasma Physics, a joint program between Princeton and PPPL. “Machine learning models can describe plasma behavior very well, and importantly, machine learning models are the only way to model plasmas in milliseconds. The speed of these models is the key to control.”

PACMAN integrates multiple AI models into one fusion control system

Machine learning has already demonstrated significant potential for controlling fusion plasmas. However, many previous applications were developed independently and lacked a common framework for allowing different models to work together. Fusion experiments require multiple AI models because different plasma conditions and machine systems must be monitored and controlled at the same time.

PACMAN was designed to provide that shared framework.

“We developed this framework to allow models to communicate, share output from those models, and do exciting physics in one integrated system,” said Andy Rothstein, a graduate student in Princeton’s Department of Mechanical and Aerospace Engineering and co-lead author of the paper.

The system combines multiple machine learning models in an iterative control loop that operates much faster than a human can.

“A really focused human operator can respond within seconds,” Rothstein said. “The entire PACMAN framework typically runs in about 20 milliseconds, but it doesn’t just run once. It runs over and over again. It can monitor small things happening inside the plasma and adjust them in ways that humans could never do.”

How PACMAN controls a fusion tokamak

PACMAN operates like an assembly line with four main stages. First, it collects real-time measurements from the tokamak, including temperature, density and magnetic signals. The system checks the data for errors and combines the measurements into a single package.

The AI model then selects the measurements it needs and uses them to estimate the plasma’s current condition or predict what may happen next. Based on these predictions, the controller determines the appropriate action, such as increasing the power of a heating beam.

In the final stage, PACMAN resolves conflicting commands, applies strict hardware safety limits and sends approved instructions to the tokamak. Because the models and controller operate independently, scientists can add new components without disrupting the rest of the framework.

PACMAN tested on a real fusion machine

Researchers demonstrated PACMAN’s flexibility in five experiments conducted on the DOE’s DIII-D National Fusion Facility tokamak in San Diego.

During these tests, PACMAN was able to:

• Use a reinforcement learning model, trained through trial and error, to take full control of the plasma heating system.
• Predict a sudden burst of energy from the edge of the plasma.
• Detect and control waves in the plasma driven by high-speed particles.
• Adjust plasma density and rotation to meet goals set by researchers.
• Predict a tearing-mode instability and prevent it before it developed.

The tearing-mode experiments highlighted one of the system’s key advantages. Traditional controllers cannot detect this instability until it has already started.

“Then they try to suppress it, but that can come with significant performance degradation,” Kaga said. “In one of the experiments we present, a machine learning model predicts tearing mode about 200 milliseconds in advance, allowing us to modify the plasma to avoid tearing mode in the first place.”

PACMAN also adjusted all six of DIII-D’s gyrotrons simultaneously. These systems heat plasma with powerful microwave beams. To achieve complex targets selected in advance by researchers, the framework changed the gyrotrons’ output while moving their mirrors in real time.

“Until now, no algorithm existed to find such an optimal solution,” Kaga said. “When we looked at the data after the shot, it was doing exactly what we wanted it to do, with all six moving at the same time in the optimal way to get to the goal.”

AI accelerates fusion experiments while keeping humans in control

Rothstein said one of the most surprising results was how quickly PACMAN enabled researchers to deploy additional AI models. Building the framework and installing the first model required several months of work.

“Then we rolled out a second model, which took a few days. Testing was easier and there were far fewer bugs,” he said. “DIII-D is first and foremost a research machine, and sometimes things don’t go as expected. If you can fit a model in one week, you can retrain and fit a new model the next week. It allows you to do iterations that weren’t possible before.”

The researchers emphasize that PACMAN is not designed to remove humans from fusion experiments. The framework enforces hardware safety limits regardless of what the AI model recommends. This allows physicists to review results after each experiment and refine the controller before the next test.

“No matter how powerful the controller is, ultimately it is the human operator who sets the parameters of that controller,” Kaga said.

A flexible AI platform for future fusion devices

PACMAN’s modular structure could also be useful beyond DIII-D. Its developers believe the framework can be adapted to tokamaks with different shapes, sizes and equipment, including future fusion machines that have not yet been designed.

“PACMAN uses a flexible setup that allows you to combine building-block AI algorithms. You can add new ones, replace one, or run several at the same time without touching the rest of the system,” said Egemen Koremen, associate professor of mechanical and aerospace engineering at Princeton University, with joint appointments at the Andlinger Center for Energy and the Environment and PPPL. “Its modularity transforms AI plasma control from a series of one-off demonstrations to an infrastructure that the entire fusion community can build upon.”

Other authors of this paper are: Ricardo Chauchat, Keith Erickson, Kim Sang Kyung, PPPL’s Jalal-ud-din Butt, Peter Steiner, Azarakhsh Jalalvand of Princeton University, and Takuma Wakatsuki of the National Institutes for Quantum Science and Technology in Japan.

This research was supported by the DOE Office of Science using the DIII-D National Fusion Facility under awards DE-FC02-04ER54698, DE-SC0015480 and DE-AC02-09CH11466, and by a National Science Foundation Graduate Research Fellowship under grant DGE-2039656.