An AI steered a fusion reactor faster than a human could react: what PPPL's PACMAN does, and what it doesn't

TL;DR

Researchers at the US Princeton Plasma Physics Laboratory have shown that an AI system called PACMAN can predict and head off a dangerous plasma instability in a tokamak fusion reactor faster than a human operator could react. In one of five live experiments on the DIII-D tokamak (its results published in the journal Nuclear Fusion), PACMAN forecast a "tearing mode" about 200 milliseconds before it would strike and adjusted the tokamak to avoid it, rather than reacting after the disruption began. It is a real advance in AI control of nuclear fusion, but it is instability control on an experimental machine, not net energy gain and not a power plant, and it does not by itself move up any commercial-fusion timeline.

"AI runs nuclear fusion" is the kind of headline that is exciting and, taken literally, wrong. What actually happened is narrower and, for anyone following fusion, more genuinely useful: an AI learned to see a reactor-wrecking instability coming and to steer around it before it hit. The system, PACMAN, comes from the US Princeton Plasma Physics Laboratory (PPPL) and was tested on a real tokamak. Here is how it works, why predicting trouble ahead of time is the hard part, and why this is a control milestone rather than an energy one.

What PACMAN actually did

A tokamak confines a superheated, electrically charged gas with powerful magnetic fields so that fusion reactions can occur. One of the things that can ruin a run is a tearing mode, an instability in which the magnetic surfaces holding the gas break up, which can abruptly end the reaction and stress the machine. The usual approach is reactive: detect the disruption and respond. PACMAN's advance is to be predictive.

PPPL reports that PACMAN was tested in five live experiments on the DIII-D tokamak, each demonstrating a different control task, from running the heating systems to predicting edge energy bursts and controlling plasma density. In one of those experiments, it:

Predicted a tearing-mode instability about 200 milliseconds before it would occur, giving the control system time to act preemptively; and
Adjusted the tokamak's actuators to steer the reaction away from the instability, avoiding the disruption rather than cleaning up after it.

The design and results were published in the peer-reviewed journal Nuclear Fusion.

That 200-millisecond figure is the prediction lead time, the warning window between the forecast and the instability itself. It is enough for PACMAN's controller, which runs its full loop in roughly 20 milliseconds, to change course, where a human operator, reacting on the order of seconds, could not.

Why predicting the instability early is the hard part

Reacting to a disruption once it has started is often too late; the reaction can be over in the time it takes to respond. Getting ahead of it means reading subtle precursors in the reactor's signals and correctly forecasting that a tearing mode is forming, then knowing which adjustments will defuse it without wrecking the conditions you need for fusion. That is a pattern-recognition-and-control problem well suited to machine learning, and it is where PACMAN's contribution sits: not generating energy, but keeping the reaction stable and alive long enough to be useful. Steady, disruption-free operation is one of the practical hurdles between today's experiments and any future power plant.

What this does NOT mean

Because the hype gap is large here, the caveats matter as much as the result:

It is not net energy gain. PACMAN is about controlling the reaction's stability, not about producing more energy than the reactor consumes. This is a separate milestone that this work does not claim.
It is not a power plant. DIII-D is an experimental research tokamak. Demonstrating instability control there does not turn on a commercial reactor.
It does not move the commercial timeline by itself. Better control is necessary but not sufficient; it is one contribution among many on a long road.
It is distinct from earlier AI-fusion work. A separate 2024 study used deep reinforcement learning to avoid tearing modes; PACMAN is a different system and result from a different effort, not the same thing rebranded.

Read plainly, this is an important step in making fusion reactors controllable enough to one day run reliably, which is exactly the kind of unglamorous problem that has to be solved before the glamorous ones. The wider race to make fusion practical is a running thread in our future coverage.

Frequently asked questions

Can AI control a nuclear fusion reactor?

In a narrow, real sense, yes: PPPL's PACMAN system predicted a plasma-destroying instability in the DIII-D tokamak about 200 milliseconds ahead and adjusted the tokamak to avoid it, in one of five live experiments. It controls the reaction's stability; it does not generate energy or run a power plant.

What is a tearing-mode instability?

It is an instability in which the magnetic surfaces confining a tokamak's superheated gas break up, which can abruptly end the fusion reaction and stress the machine. Predicting and avoiding tearing modes is one of the key control problems in fusion research.

Does this mean fusion power is closer?

Not directly. PACMAN is a control advance, not an energy one, and it does not by itself move up any commercial-fusion timeline. Reliable, disruption-free operation is one of many hurdles between experiments and a working power plant.

What is a tokamak?

A tokamak is a doughnut-shaped device that uses strong magnetic fields to confine a superheated, electrically charged gas so that fusion reactions can occur. DIII-D, where PACMAN was tested, is an experimental research tokamak in the United States.

What is PACMAN, and who built it?

PACMAN is an AI framework built by the US Princeton Plasma Physics Laboratory to predict and preemptively avoid plasma instabilities in fusion reactors. It was tested on the DIII-D tokamak and the results were published in the journal Nuclear Fusion.