An artist’s interpretation of the PACMAN artificial intelligence framework for fusion systems. Credit: Kyle Palmer / PPPL Communications Department
Princeton’s PACMAN AI can control fusion plasma in milliseconds and predict dangerous instabilities before they start.
In some fusion energy systems, particles can reach temperatures hotter than the center of the Sun. The challenge is keeping that extreme plasma under control, because disturbances can develop within just a few thousandths of a second, much faster than a person could respond.
Researchers at the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL) and Princeton University have developed a new software framework that uses artificial intelligence (AI) to make those rapid control decisions. The system is designed to respond at machine speed while maintaining strict safety protections and leaving the overall goals in human hands.
The framework is called PACMAN (a novel abbreviation for Prediction And Control using MAchiNe learning). Researchers successfully tested it in five experiments on a working fusion system. Its design and initial results are described in a new paper published in Nuclear Fusion.
Keeping Fusion Plasma Stable
Fusion has the potential to provide a virtually unlimited source of electricity. Scientists are exploring several ways to reproduce the process on Earth, including machines called tokamaks. These devices use strong magnetic fields to confine a plasma: an electrically charged gas often described as the fourth state of matter.
Maintaining a plasma that is sufficiently hot, dense, and stable requires continual adjustments to a tokamak’s heating systems, magnets, and gas injectors. Even relatively small disturbances, known as instabilities, can grow within milliseconds and interfere with the fusion reaction.
Predicting plasma behavior is especially difficult. Advanced computer simulations can require days or even months to calculate what the plasma will do. While those simulations are valuable for planning experiments, they are far too slow to guide an experiment in real time when the entire run may last only minutes.
“That’s great for preparing for the next experiment in a year, but for control we need models that make a decision in the moment,” said co-lead author Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics, which is a joint program of Princeton University and PPPL. “Machine learning models can describe the plasma behavior very well, and importantly, they are the only way we have to model the plasma in millisecond times. The speed of these models is what’s key for control.”
Bringing Multiple AI Models Together
Machine learning has already shown considerable potential for controlling fusion plasma. However, many previous efforts were created individually from the ground up, without a shared framework that made it easy for different models to work together. That presents a problem because controlling a fusion system requires monitoring and managing many different plasma behaviors at once.
PACMAN was built to provide that common structure.
“We developed this framework so that models could communicate, outputs from those models could be shared, and we could do exciting physics in one integrated system,” said Andy Rothstein, a graduate student at Princeton University’s Department of Mechanical and Aerospace Engineering and co-lead author of the paper.
PACMAN combines several machine learning models into a continuous control loop that operates far faster than a human could.
“A really focused human operator can respond on the order of seconds,” Rothstein said. “The whole PACMAN framework typically runs in about 20 milliseconds, and it’s not running once. It’s running again and again and again. It can see small things happening in the plasma and adjust in a way that a human would never be able to do.”
The system functions much like an assembly line divided into four stages. It begins by collecting measurements from the tokamak in real time, including temperature, density and magnetic signals. Those measurements are checked for errors and organized into a single package.
AI models then select the measurements they need and estimate what the plasma is currently doing or what it is likely to do next. Controllers use those predictions to determine appropriate actions, such as increasing the strength of a heating beam. In the final stage, the system resolves any conflicting instructions, applies strict hardware safety limits, and sends the approved commands to the tokamak.
Because the individual models and controllers operate independently, researchers can introduce new components without rebuilding or disrupting the entire system.
AI Tested on a Real Fusion Machine
Researchers demonstrated PACMAN’s flexibility during five experiments at the DOE’s DIII-D National Fusion Facility tokamak in San Diego.
During those tests, PACMAN:
Gave complete control of the heating systems to an AI model trained using a trial-and-error method known as reinforcement learning.Predicted sudden energy bursts emerging from the edge of the plasma.Identified and controlled plasma waves caused by fast-moving particles.Adjusted plasma density and rotation to targets selected by the researchers.Predicted an instability known as a tearing mode and prevented it from developing.
The tearing mode experiment was particularly notable. Traditional controllers can only identify this type of instability once it has already begun.
“Then they try to suppress it, and that can come with a lot of performance degradation,” Farre Kaga said. “In one of the experiments we present, a machine learning model predicts the tearing mode about 200 milliseconds in advance, so the plasma can be changed to avoid it in the first place.”
Coordinating Six Plasma Heating Systems at Once
PACMAN also controlled all six of DIII-D’s gyrotrons (systems that heat the plasma with powerful microwave beams) at the same time.
To accomplish complex goals established beforehand by the researchers, the framework continuously changed the direction of the gyrotron mirrors and adjusted their power levels as the experiment was running.
“There was no algorithm to find that optimal solution before,” Farre Kaga said. “When the shot ended, and we looked at the data, it was doing exactly what we hoped, simultaneously moving all six in an optimal way to reach the goal.”
Making Fusion AI Faster to Develop and Test
Another advantage of PACMAN is the speed at which researchers can integrate new machine learning models.
Rothstein said creating the framework and installing the first model required months of work. Adding the second model was dramatically faster.
“Then we went to put in the second model, and it took a couple of days. The 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 work out the way you expected. If you can put a model on in a week, you can retrain it and put a new one on the week after. It allows for iteration that wasn’t possible previously.”
That ability to rapidly introduce, test, and improve models could help fusion researchers experiment with new approaches much more quickly.
Humans Remain in Control
Despite giving AI responsibility for extremely fast decisions, the researchers emphasize that PACMAN does not remove humans from the process.
The framework applies hardware safety limits regardless of what an individual AI model recommends. Physicists also examine each experiment afterward and adjust the controllers before subsequent runs.
“No matter how sophisticated your controllers, in the end it’s a human operator that sets the parameters for that control,” Farre Kaga said.
PACMAN’s modular design could also allow it to be used beyond DIII-D. Its developers believe the same approach could be adapted to tokamaks with different dimensions, shapes and instruments, including future fusion machines that have not yet been designed.
“PACMAN uses a flexible setup where building-block AI algorithms can be put together. You can add a new one, swap one out, or run several at once without touching the rest of the system,” said Egemen Kolemen, associate professor of mechanical and aerospace engineering at Princeton University, jointly appointed with the Andlinger Center for Energy and the Environment and PPPL. “That modularity is what turns AI plasma control from a series of one-off demonstrations into infrastructure the whole fusion community can build on.”
Reference: “Enabling integrated AI control on DIII-D: a control system design with state-of-the-art experiments” by A. Rothstein, H.J. Farre-Kaga, J. Butt, R. Shousha, K. Erickson, T. Wakatsuki, P. Steiner, S.K. Kim, A. Jalalvand and E. Kolemen, 2 July 2026, Nuclear Fusion.
DOI: 10.1088/1741-4326/ae7f9d
Other authors on the paper include Ricardo Shousha, Keith Erickson and SangKyeun Kim from PPPL, Jalal-ud-din Butt, Peter Steiner and Azarakhsh Jalalvand from Princeton University, and Takuma Wakatsuki from Japan’s National Institutes for Quantum Science and Technology.
The 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 the National Science Foundation Graduate Research Fellowship under grant DGE-2039656.
Never miss a breakthrough: Join the SciTechDaily newsletter.
Follow us on Google and Google News.