Princeton AI Tames Fusion Plasma Hotter Than the Sun

Princeton University researchers have created an AI system called PACMAN. This system can control fusion plasma that is hotter than the sun. It can also predict dangerous instabilities before they happen.

Fusion energy systems aim to create clean electricity. They involve heating particles to extreme temperatures, sometimes hotter than the sun's core. Keeping this super-hot plasma stable is a major challenge. Disturbances can occur in milliseconds, too fast for humans to react.

PACMAN stands for Prediction And Control using MAchiNe learning. It's a new software framework developed by the U.S. Department of Energy’s (DOE) Princeton Plasma Physics Laboratory (PPPL) and Princeton University. This AI makes rapid control decisions at machine speed. It also includes strict safety features, keeping humans in charge of the overall goals.

Researchers successfully tested PACMAN in five experiments on a working fusion system. Their findings were published in Nuclear Fusion.

Keeping Fusion Plasma Stable

Fusion could provide an almost endless source of electricity. Scientists are trying to replicate this process on Earth using machines called tokamaks. These devices use strong magnetic fields to hold plasma, which is an electrically charged gas.

To keep plasma hot, dense, and stable, a tokamak's heating systems, magnets, and gas injectors need constant adjustments. Even small disturbances, called instabilities, can grow in milliseconds and disrupt the fusion reaction.

Predicting plasma behavior is very hard. Advanced computer simulations can take days or months to calculate what the plasma will do. These simulations are useful for planning experiments, but they are too slow for real-time control during an experiment that might only last minutes.

Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics, explained that machine learning models can describe plasma behavior well. More importantly, they are the only way to model plasma in milliseconds, which is crucial for control.

Combining Multiple AI Models

Machine learning has shown great promise for controlling fusion plasma. However, many past efforts built models individually. This made it hard for different models to work together. Controlling a fusion system requires monitoring many plasma behaviors at once.

PACMAN was designed to provide a common structure for these models.

Andy Rothstein, a graduate student at Princeton University, said the framework allows models to communicate and share outputs. This enables exciting physics research within one integrated system.

PACMAN combines several machine learning models into a continuous control loop. This loop runs much faster than a human could.

Rothstein noted that a focused human operator can respond in seconds. The PACMAN framework typically runs in about 20 milliseconds, and it does so repeatedly. It can detect small changes in the plasma and adjust in ways a human cannot.

The system works like an assembly line with four stages. First, it collects real-time measurements from the tokamak, such as temperature, density, and magnetic signals. These measurements are checked for errors and organized.

Next, AI models use these measurements to estimate what the plasma is doing or will do. Controllers then use these predictions to decide on actions, like increasing a heating beam's strength. Finally, the system resolves any conflicting instructions, applies safety limits, and sends commands to the tokamak.

Because the individual models and controllers work independently, researchers can add new parts without rebuilding the whole system.

AI Tested on a Real Fusion Machine

Researchers showed PACMAN's flexibility during five experiments at the DOE’s DIII-D National Fusion Facility tokamak in San Diego.

During these tests, PACMAN:

• Gave full control of heating systems to an AI model trained with a trial-and-error method called reinforcement learning.
• Predicted sudden energy bursts from the plasma's edge.
• Identified and controlled plasma waves caused by fast-moving particles.
• Adjusted plasma density and rotation to targets set by researchers.
• Predicted and prevented an instability called a tearing mode.

The tearing mode experiment was especially important. Traditional controllers can only spot this instability after it has started.

Farre Kaga explained that traditional methods try to suppress the instability, which can hurt performance. In one experiment, a machine learning model predicted the tearing mode about 200 milliseconds ahead of time. This allowed the plasma to be changed to avoid the instability entirely.

Coordinating Six Plasma Heating Systems

PACMAN also controlled all six of DIII-D’s gyrotrons at once. Gyrotrons heat the plasma with powerful microwave beams.

To achieve complex goals set by researchers, the framework continuously changed the direction of the gyrotron mirrors and adjusted their power levels during the experiment.

Farre Kaga said there was no algorithm to find this optimal solution before. When they reviewed the data, PACMAN was doing exactly what they hoped, moving all six gyrotrons optimally to reach the goal.

Faster Fusion AI Development

Another benefit of PACMAN is how quickly researchers can add new machine learning models.

Rothstein said creating the framework and installing the first model took months. Adding the second model was much faster, taking only a couple of days. Testing was easier, and there were fewer bugs. This allows for faster iteration, which was not possible before.

This ability to quickly introduce, test, and improve models could help fusion researchers experiment with new approaches much faster.

Humans Remain in Control

Even though AI makes very fast decisions, humans still control the process.

The framework applies hardware safety limits no matter what an individual AI model suggests. Physicists also review each experiment and adjust the controllers before future runs.

Farre Kaga emphasized that human operators still set the parameters for control, no matter how advanced the controllers are.

PACMAN's design also means it could be used beyond DIII-D. Its developers believe the same approach could work for tokamaks of different sizes, shapes, and instruments, including future fusion machines.

Egemen Kolemen, a professor at Princeton University and PPPL, noted that PACMAN uses a flexible setup where AI algorithms can be combined. You can add, swap, or run several without affecting the rest of the system. This modularity turns AI plasma control into infrastructure the entire fusion community can build on.

Deep Dive & References

• Enabling integrated AI control on DIII-D: a control system design with state-of-the-art experiments - Nuclear Fusion, 2026