Software built at the U.S. Department of Energy's Princeton Plasma Physics Laboratory and Princeton University ran real-time control of fusion plasma in five experiments at the DIII-D National Fusion Facility in San Diego, according to a paper in Nuclear Fusion. In one test, a machine learning model predicted a tearing mode instability about 200 milliseconds before it appeared, and the plasma was adjusted to avoid it.

The framework is called PACMAN, short for Prediction And Control using MAchiNe learning. Its full cycle typically runs in about 20 milliseconds and repeats continuously. Co-lead author Andy Rothstein noted that even a very focused human operator reacts on the order of seconds.

That speed gap explains why some of fusion's hardest near-term problems look like control problems, where milliseconds matter. The result is not net energy gain, not a power plant and not electricity on anyone's grid. It is a research tool demonstrated on a research machine.

A Four-Stage Loop Running in Milliseconds

Tokamaks use powerful magnetic fields to confine plasma, an electrically charged gas, in a doughnut-shaped chamber, with particles that can be hotter than the sun's core. The plasma must stay hot, dense, and stable, and small disturbances called instabilities can grow within milliseconds. Detailed physics simulations can take days or even months, far too slow to guide a live experiment.

PACMAN works like a four-station assembly line, according to Princeton. It gathers live temperature, density, and magnetic measurements and checks them for errors. Machine learning models then estimate what the plasma is doing or is likely to do next. Controllers turn those estimates into actions, such as raising the power of a heating beam. A final stage resolves conflicting commands, applies strict hardware safety limits and sends approved instructions to the machine.

Across the five experiments, the system let a reinforcement learning model take complete control of the heating systems, predicted sudden bursts of energy from the plasma's edge, detected and controlled waves driven by fast particles, adjusted density and rotation to researcher-set targets, and headed off a tearing mode. It also coordinated all six of DIII-D's gyrotrons, which heat the plasma with microwave beams, adjusting their power and repositioning their mirrors in real time. The team's preprint describes cycle-time goals of roughly 5 to 50 milliseconds, depending on the task.

Co-lead author Hiro Farre Kaga said physics simulations are too slow for live decisions. "The speed of these models is what's key for control," he said.

The Value of a 200-Millisecond Warning

A tearing mode occurs when magnetic field lines inside the plasma break, creating an opening for the plasma to escape and potentially ending the reaction. Conventional controllers cannot identify this instability until it has begun, Farre Kaga explained, and suppressing it at that point can come with a significant loss of performance.

A 200-millisecond warning is shorter than a slow blink. For a loop running every 20 milliseconds, however, it amounts to about 10 decision cycles before trouble begins. Other instabilities matter too: the Department of Energy has highlighted research on suppressing large, damaging edge energy bursts as a way to protect future fusion devices, and edge-burst prediction was one of PACMAN's five tests.

Rothstein said one of the most surprising results was how quickly new models could be added. The first model took months to install, but the second took a couple of days, which lets researchers retrain and redeploy models from one week to the next.

An Evidence Check on the Fusion AI Claim

This is not the first time AI has predicted this instability. In 2024, researchers led by the same Princeton group reported in Nature that a model could forecast tearing modes up to 300 milliseconds ahead on DIII-D. The new contribution is an integrated framework that lets many models and controllers run together safely, which principal investigator Egemen Kolemen described as turning one-off demonstrations into shared infrastructure.

The study is a peer-reviewed engineering and experimental paper covering five experiments on one device. A preprint version appeared in November 2025. The developers believe the modular design could be adapted to tokamaks of other shapes and sizes, but that has not yet been tested. The paper does not claim any fusion energy output, and Rothstein described DIII-D as first and foremost a research machine.

Humans remain in charge. Researchers set the objectives, hardware safety limits apply no matter what a model recommends, and physicists review results after each experiment. The work was supported by the DOE Office of Science and a National Science Foundation Graduate Research Fellowship.

A Long Road from Lab Control to the Power Grid

The DOE Fusion Energy Sciences program describes fusion as a possible long-term energy source that uses abundant fuel and does not produce long-lived radioactive waste. It also says foundational science and technology gaps must be resolved before fusion can become an affordable and reliable energy source. Artificial intelligence and machine learning are among the areas the program funds.

For households, nothing about utility bills or power supply changes because of this result. Fusion is not supplying electricity to the U.S. grid today. Readers weighing future breakthrough headlines can look for concrete milestones, such as sustained net energy, long-duration operation, and actual power plant construction.

The team's next goals, described in the preprint, include running models and controllers in parallel to cut computing time and exploring faster hardware such as GPUs or FPGAs. Nature World News will follow as those upgrades are tested.

AI can now manage several plasma control tasks, faster than any person, on a working U.S. fusion device. That is a meaningful engineering step. The central uncertainty is whether the approach scales to the larger, hotter machines that commercial fusion would require.

What Readers Want to Know

What did the Princeton team demonstrate? An AI framework, PACMAN, controlled several aspects of fusion plasma in real time during five experiments at the DIII-D tokamak in San Diego.

What is a tearing mode? It is a plasma instability in which magnetic field lines break, which can let the plasma escape and end the reaction.

Is this a fusion energy breakthrough? No. The work improves control of experiments. It does not produce net energy or electricity.

Is AI running the machine without people? No. Researchers set the goals, hardware safety limits always apply, and physicists review each experiment.

Was this the first AI prediction of a tearing mode? No. A 2024 Nature study by the same research group forecast tearing modes up to 300 milliseconds ahead.