← Back to BracksterNews Close-up of a glowing plasma globe with vibrant electric sparks against a dark background, illustrative photo

World - Science & Health

Princeton Built an AI That Reacts Inside a Fusion Reactor Faster Than Any Human Ever Could

Inside a working tokamak, plasma hotter than the sun's core can spiral out of control in a few thousandths of a second — far too fast for any operator watching a screen. A new AI framework out of Princeton Plasma Physics Laboratory now reads that plasma and issues corrections every 20 milliseconds, and it just passed its first five real tests on an actual fusion machine.

By Gustavo Henrique · About the author

PRINCETON, N.J. · SEP 2-6 ~6 MIN READ

Updated on September 8, 2026 · Editorial update: source review, context and clarity.

Fusion energy has always had a timing problem. Getting a plasma hot and dense enough to fuse atoms is one challenge; keeping it stable long enough to matter is another one entirely, because the disturbances that ruin an otherwise perfect run — physicists call them instabilities — can build and spread within a few milliseconds. A human operator, no matter how well trained, reacts on the order of seconds. That gap between how fast plasma misbehaves and how fast people can respond has quietly limited fusion research for decades. This month, researchers at the Department of Energy's Princeton Plasma Physics Laboratory (PPPL) and Princeton University said they've built something that closes it.

Summary

Inside a working tokamak, plasma hotter than the sun's core can spiral out of control in a few thousandths of a second — far too fast for any operator watching a screen.
A new AI framework out of Princeton Plasma Physics Laboratory now reads that plasma and issues corrections every 20 milliseconds, and it just passed its first five real tests on an actual fusion machine.
Fusion energy has always had a timing problem.

An AI framework named after a very different kind of pattern-chasing

The system is called PACMAN, short for Prediction And Control using MAchiNe learning, and its job is almost comically literal: chase down small problems in the plasma before they become big ones. Rather than relying on a single purpose-built model, PACMAN is designed as a framework that lets multiple AI models plug directly into a fusion experiment's control system, reading live measurements of temperature, density, and magnetic fields, then issuing corrective commands automatically — all while keeping human researchers in charge of the actual experimental goals. The team detailed the framework's design, along with its first results, in a peer-reviewed paper published in the journal Nuclear Fusion.

Twenty milliseconds, fifty times a second, for the length of the run

The numbers are the whole story here. According to researchers involved in the project, a focused human operator can respond to a developing problem in, at best, a couple of seconds. PACMAN completes a full read-decide-correct cycle in about 20 milliseconds — and it doesn't do that once. It does it roughly fifty times every second, continuously, for as long as the experiment runs. In one test, the system reportedly predicted a damaging plasma instability a full 200 milliseconds before it actually appeared, and adjusted conditions early enough to keep it from forming at all — catching a problem before it existed, rather than reacting once it did.

A really focused human operator can respond on the order of seconds.— on why fusion control needs an AI, not a person, at the controls

Why the old approach couldn't keep up

The obvious question is why this wasn't done sooner, and the answer comes down to speed of a different kind. The detailed physics simulations researchers normally use to model plasma behavior are extremely accurate, but they can take days or even months of computing time to run — perfectly fine for planning next year's experiment, useless for making a decision inside one that might only last a few minutes. Hiro Farre Kaga, a graduate student in the joint Princeton–PPPL plasma physics program and a co-lead author on the paper, said the field needed models that could describe plasma behavior accurately while still rendering a decision in the moment the experiment is happening, not weeks after it ends. Machine-learning models, trained ahead of time on that same slow physics, turn out to be the only tool fast enough to do both at once.

Tested on a real machine, not a simulation

PACMAN wasn't just built and published — it was tried. Researchers ran the framework across five separate live experiments on DIII-D, the Department of Energy's tokamak facility in San Diego, California, one of the country's primary testbeds for magnetic confinement fusion. Each run used real plasma, real magnetic fields, and real instrumentation, with PACMAN making live control decisions rather than being evaluated after the fact against recorded data. That distinction matters in fusion research, where plenty of promising machine-learning ideas have looked strong on paper only to struggle once they're handed the controls of an actual reactor running in real time.

None of this means fusion power is suddenly close to your electrical outlet. PACMAN solves a control problem, not the far bigger challenge of getting a fusion reaction to produce more energy than it consumes at commercial scale — a milestone the field is still chasing across multiple competing reactor designs. What it does solve is one of the quieter obstacles standing between today's experimental tokamaks and the steadier, longer, more stable plasma runs that future fusion power plants will need to survive on.

This account draws on reporting and research summaries from Princeton Plasma Physics Laboratory, ScienceDaily, Newswise, Mirage News, and JFeed.

Analysis by Gustavo Henrique

Nobody's claiming PACMAN “solves” fusion, and to their credit, the researchers themselves are careful to frame this as a control-engineering advance, not a headline about limitless clean energy arriving next year. That restraint, more than the millisecond figure, is why I trust this one more than most AI-and-energy stories that cross my desk. Fusion's hardest problems are still unsolved. But every field eventually gets bottlenecked by something boring and specific — in this case, plasma moving faster than a person can watch a dial and turn a knob — and it's genuinely satisfying to see a team go solve that particular, narrow, unglamorous problem instead of just promising the moon.

BRACKSTERNEWS STANDARDS

Facts in this story were verified independently before publication; analysis is clearly labeled and kept apart from reporting. See our editorial methodology for specifics.