An AI now steers fusion plasma faster than any human could
What happened
Researchers have built an artificial intelligence system that can steer the superheated gas inside a fusion reactor faster than a human operator could ever react. The system is called PACMAN, which stands for Prediction And Control using MAchiNe learning. The work comes from the Princeton Plasma Physics Laboratory and Princeton University, with collaborators at Japan's National Institutes for Quantum Science and Technology. It was published in the journal Nuclear Fusion and reported on 6 September 2026. Here is the problem it solves. Fusion works by heating a gas until it becomes a plasma, hotter than the centre of the Sun, and holding it in place with powerful magnets. Plasma is unruly. It develops instabilities, wobbles and tears that can wreck a run in a fraction of a second. Humans cannot respond quickly enough. Our reaction times are measured in whole seconds. The plasma does its damage in milliseconds. PACMAN watches the plasma and predicts trouble before it arrives. In one test it spotted a 'tearing mode' instability roughly 200 milliseconds before it would have happened, and adjusted the machine to prevent it. Its full control cycle takes about 20 milliseconds. It also coordinated six gyrotrons at once. Gyrotrons are the devices that beam microwave energy into the plasma to heat and steady it. Getting six of them to work together towards a complex target is not something a person can do by hand. The work was co-led by Hiro Farre Kaga and Andy Rothstein, with Egemen Kolemen, an associate professor at Princeton, among the senior researchers. One of the team put the reasoning plainly: machine learning models can describe plasma behaviour very well, and they are the only way we have to model the plasma in millisecond times.
Why this matters
Fusion is the reaction that powers the Sun. If we can make it work reliably on Earth, it offers a source of electricity with no carbon emissions, no risk of a runaway meltdown, and fuel that can be drawn from water. That is a very big prize, and it has been out of reach for a long time. The running joke is that fusion is always thirty years away. But the reasons it is hard have changed. The physics is far better understood than it was. Increasingly, the obstacles are engineering and control: keeping a plasma stable, hour after hour, without something going wrong. That is precisely the gap this kind of work fills. A future power station cannot have a room of people watching dials and hoping. It needs to run itself, safely and boringly, the way a modern aircraft or a gas turbine does. There is also a wider lesson here about where AI is genuinely useful. Much of the public conversation about artificial intelligence is about machines writing text or making pictures. This is a different and, arguably, more solid use: a narrow, well-defined job, with a clear measure of success, where the machine is simply faster than a person and the stakes of being slow are high.
What the evidence actually says
It is worth being precise about what has and has not been shown. This is a control system demonstrated on experimental fusion machines. It is not a power station. No fusion plant anywhere is yet putting electricity into a grid, and none will for some years. What the research shows is that a machine learning system can predict a specific type of instability in advance and act to avoid it, within a control loop fast enough to matter. That is a concrete, measurable result, published in a peer-reviewed journal. It does not show that every kind of plasma problem can be predicted this way. Tearing modes are one important failure mode among several. There is also an important design choice worth noticing. Human operators keep final authority over the system's objectives and its safety limits. The AI is not deciding what the machine should be trying to do. It is deciding, very fast, how to get there within boundaries people have set. That is a sensible pattern, and it is one we are likely to see more of: the machine handles the millisecond decisions, the people handle the goals and the limits.
Practical advice
This is not a story with an action list attached, so here is something more useful: how to read fusion news without being misled. **Ask what was actually achieved.** 'Breakthrough' is doing a lot of work in most fusion headlines. Look for the specific claim. Here it is: an instability was predicted 200 milliseconds ahead and prevented, in a 20-millisecond control loop. **Watch for the difference between energy in and energy out.** Some fusion results are about the plasma producing more energy than was pumped into it. Others, like this one, are about control and stability. Both matter. They are not the same claim. **Check whether it is peer reviewed.** This work is in Nuclear Fusion, an established journal. That does not make it beyond question, but it does mean other physicists have gone through it. **Be patient with timelines.** Progress in fusion is real and it has genuinely sped up. It is still measured in years, not months. **Notice who keeps control.** In any story about AI running something important, one of the most useful questions is what the humans still decide. Here, the answer is clear and reassuring.
What to know
Princeton researchers have shown that an AI system, PACMAN, can predict and prevent a damaging plasma instability inside a fusion machine, working roughly a hundred times faster than a human operator. It predicted one instability about 200 milliseconds before it would have occurred, completed its control decisions in around 20 milliseconds, and coordinated six heating devices at once. Humans keep control of the goals and the safety limits. The AI handles the split-second adjustments. This does not mean fusion power is imminent. It does mean one of the practical obstacles between a laboratory experiment and a machine that could run continuously has moved a little closer to being solved. Sources: ScienceDaily, 'AI can now control fusion plasma faster than humans can react', 6 September 2026, https://www.sciencedaily.com/releases/2026/09/260903064215.htm | Nuclear Fusion, 2026, volume 66, article 076050 (Princeton Plasma Physics Laboratory), https://iopscience.iop.org/journal/0029-5515 | Princeton Plasma Physics Laboratory, https://www.pppl.gov/ This article is for general information. It is about energy research and contains no medical advice; for anything concerning your health, please speak to a doctor, pharmacist or other qualified healthcare professional.
Related articles
A knitted fabric that snaps into shape - and works as a switch
Engineers at Harvard have made knitted fabrics that can snap between different shapes and hold them, and that can double as soft electronic switches. The resea…
Scientists have turned bacteria into living transistors
Engineers at MIT have made bacteria behave like the components inside a computer. The work was published in Nature Chemical Biology and reported on 4 September…