Plate Nº 51 · recorded October 10, 2026
PhysicsReported finding
Princeton AI Predicts Fusion Plasma Instabilities 200 Milliseconds Early
A Princeton AI framework called PACMAN runs fusion plasma control in 20-millisecond loops and predicted a damaging tearing mode instability 200 milliseconds before it appeared across five tests on the DIII-D tokamak in San Diego.
By Marcus Bennett4 min read831 words
In brief
- The PACMAN framework runs its control loop in about 20 milliseconds, while a focused human operator responds on the order of seconds.
- In one DIII-D experiment, a PACMAN model predicted a tearing mode instability about 200 milliseconds before it formed.
- Researchers tested the system in five experiments on the DIII-D National Fusion Facility in San Diego.
- PACMAN simultaneously coordinated all six DIII-D gyrotrons, adjusting their power and mirror angles in real time.
- The work was published September 6, 2026, in the journal Nuclear Fusion (DOI: 10.1088/1741-4326/ae7f9d).
A new artificial intelligence framework tested at Princeton runs its control loop in about 20 milliseconds — far faster than the seconds a skilled human operator needs. In five experiments on a working fusion machine, the system also predicted a damaging plasma instability roughly 200 milliseconds before it appeared.
Researchers at the U.S. Department of Energy's Princeton Plasma Physics Laboratory (PPPL) and Princeton University built the framework, called PACMAN (short for Prediction And Control using MAchiNe learning). They describe the design and initial tests in a paper published September 6, 2026, in the journal Nuclear Fusion.
What problem does PACMAN solve?
Fusion reactors make energy by heating a gas — known as plasma, the electrically charged "fourth state of matter" — to temperatures hotter than the sun's core. In machines called tokamaks, powerful magnets hold that plasma in place. Even small disturbances can grow in milliseconds and shut the reaction down.
Human reflexes are too slow to correct them, and traditional simulations take days to run. "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.
Machine learning models can run at millisecond speeds. Earlier fusion-AI work, however, was piecemeal — each model built separately, with no shared way to coordinate. Tokamaks need many models running at once because temperature, density, magnetic fields, and heating systems all require simultaneous control.
How does PACMAN control a tokamak?
PACMAN works like a four-step assembly line. It pulls live measurements from the tokamak — temperature, density, magnetic signals. It checks those readings for errors, then groups them into a single package.
Machine learning models then pick the data they need and estimate what the plasma is doing now or what it will do next. Controllers turn those predictions into actions, such as raising the power of a heating beam. PACMAN resolves conflicts between controllers, enforces hardware safety limits, and sends the final commands to the device.
"A really focused human operator can respond on the order of seconds," said co-lead author Andy Rothstein, a graduate student in mechanical and aerospace engineering at Princeton. "The whole PACMAN framework typically runs in about 20 milliseconds, and it's not running once. It's running again and again and again."
What did the experiments show?
Researchers tested PACMAN on the DOE's DIII-D National Fusion Facility, a tokamak in San Diego. Across five experiments, the framework:
- Let a reinforcement-learning AI take full control of the heating systems.
- Predicted sudden energy bursts at the plasma's edge.
- Detected and controlled waves driven by fast particles.
- Adjusted plasma density and rotation toward researcher-set targets.
- Predicted a "tearing mode" instability and stopped it before it formed.
The tearing mode test stood out. Conventional controllers spot this instability only after it has started, when suppressing it costs significant reactor performance. "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," Farre Kaga said.
PACMAN also coordinated all six of DIII-D's gyrotrons — devices that heat plasma with microwave beams — adjusting their power and mirror angles in real time. No algorithm had previously optimized those six beams together.
Does PACMAN replace human operators?
No. The framework enforces hardware safety limits regardless of what any AI recommends. Physicists still inspect results after each test and refine the controllers. "No matter how sophisticated your controllers, in the end it's a human operator that sets the parameters for that control," Farre Kaga said.
Rothstein added that PACMAN also speeds up research itself. Installing the first model took months. Adding the second took a couple of days, with far fewer bugs. That faster turnaround lets teams retrain models weekly — iteration not previously possible on a research tokamak.
Could PACMAN work on other fusion machines?
The framework is built from interchangeable parts. "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, jointly appointed with the Andlinger Center for Energy and the Environment and PPPL.
That modularity, the team argues, makes PACMAN a candidate for tokamaks of different shapes and sizes, including fusion machines that have not yet been built.
The study remains preliminary: five test runs on a single device. The team has not yet demonstrated the system on a larger or burning-plasma machine, and open questions remain about long-term model reliability and integration with future reactors.
Other authors include Ricardo Shousha, Keith Erickson, and SangKyeun Kim at PPPL; Jalal-ud-din Butt, Peter Steiner, and Azarakhsh Jalalvand at Princeton; and Takuma Wakatsuki at Japan's National Institutes for Quantum Science and Technology. Funding came from the DOE Office of Science and an NSF Graduate Research Fellowship.
via pppl.gov (Original)
More from Marcus Bennett
Nearby plates
- AI Scientist Makes Biological Discoveries With Minimal Human Help
- Vienna hosts world's first self-stabilizing nuclear clock
- Edinburgh Team Cuts Magnetic Memory Energy Use by Orders of Magnitude
- Physicists bridge ultraslow spin glasses and ultrafast black-hole math
- Caltech Team Directly Measures Quantum Energy Ladder Predicted for 40 Years