← All articles

⚡ Energy

20 Milliseconds: The AI That Sees a Fusion Meltdown 200 Milliseconds Before It Forms

On the DIII-D tokamak, Princeton's PACMAN framework ran a full AI control loop every 20 milliseconds, spotted a tearing-mode instability 200 milliseconds before it existed, and reshaped the plasma so it never formed. A human operator, who reacts in seconds, would have watched it happen.

By Muse · Energy · · ☕ 6 min read

Twenty milliseconds. That is how long the whole thing takes: read every sensor on the tokamak, check the data, ask the AI models what the plasma is about to do, compute the commands, resolve conflicts, enforce the hard safety limits, and send the orders back to the machine, then repeat, fifty times a second.

Princeton Plasma Physics Laboratory and Princeton University researchers built the framework, called PACMAN (Prediction And Control using MAchiNe learning), and tested it in five experiments on the DIII-D National Fusion Facility in San Diego, with the design and first results appearing in Nuclear Fusion this month. It is the first integrated, modular AI control loop running closed-loop on a real tokamak: humans set the goals and review each shot; the models take the millisecond decisions.

Inside a tokamak, the problem is speed, not intelligence. Plasma hotter than the core of the sun can go unruly in a few thousandths of a second. A focused human operator reacts in seconds: not even late, just irrelevant. Conventional controllers react after an instability has started and try to suppress it, at the cost of performance, while full-physics simulations take days or months to run, which Rothstein's team notes is "great for preparing for the next experiment in a year" and useless for control in the moment.

PACMAN works like an assembly line with four stations: first it gathers live measurements of temperatures, densities, and magnetic signals, then independent AI models predict what the plasma is doing or about to do, and the line keeps moving. Controllers turn those predictions into commands, like raising a heating beam's power. A final stage settles disagreements, applies strict hardware safety limits, and sends approved commands to the tokamak. Each model runs independently, so researchers can add, swap, or stack models without rebuilding the system.

The 10-cycle warning budget

In one experiment, an ML model predicted a tearing mode, a magnetic island conventional controllers only detect after onset, 200 milliseconds in advance, which divides by the loop time into ten full sense-predict-command cycles to act before the instability even forms.

Warning budget: control cycles before a 200 ms instability
ControllerResponse timeCycles before a 200 ms instability
PACMAN (20 ms loop)20 ms10
Focused human operator~2 s0
Conventional tearing controllerafter onsetnegative (it reacts late)

Ten cycles versus zero. That is the whole paper in one fraction. A human gets ten looks at a problem that has not happened yet; the operator gets a postmortem. Disruption prediction alone is older news: a 2019 Nature paper showed deep learning predicting disruptions 500 to 700 milliseconds out on DIII-D, but PACMAN's advance is not the alarm, it is the closed loop, prediction wired directly to action across heating, density, rotation, and stability at once.

Twelve coupled knobs

During the five experiments, a reinforcement learning model trained by trial and error rather than fed physics equations took complete control of the heating systems, while other models predicted edge energy bursts, detected fast-particle-driven plasma waves, and held density and rotation at researcher-set targets. Most striking, the framework simultaneously steered all six of DIII-D's gyrotrons, retargeting their microwave-beam mirrors and adjusting power in real time: six gyrotrons, each with a mirror aim and a power level, make twelve control variables optimized jointly every 20 milliseconds.

Farre Kaga's line: "There was no algorithm to find that optimal solution before." A human tunes knobs serially, one at a time, over seconds. Joint optimization over twelve variables in 20 milliseconds is not a faster human but a different species of control.

The 45x iteration speedup

Rothstein's biggest surprise was not the plasma but the plumbing. Building PACMAN and installing the first model took months while the second took "a couple of days," with easier testing and far fewer bugs, so call months a conservative 90 days against 2 days: roughly a 45-fold acceleration in the model deployment cycle. On a research machine where shots often misbehave, swapping a retrained model in a week instead of a quarter changes what experiments are thinkable.

Kolemen frames the point as infrastructure: modularity "turns AI plasma control from a series of one-off demonstrations into infrastructure the whole fusion community can build on." Earlier efforts were built from scratch each time, while PACMAN is the first attempt at the operating-system layer: swappable models, framework-enforced safety limits, humans on parameters and review.

Why avoidance beats suppression at ITER scale

The commercial argument lives one size class up, at ITER, where a disruption's thermal quench dumps an estimated 120 to 175 megajoules, and runaway electrons can deposit tens of megajoules onto centimeter-scale impact sites. That is why ITER treats every disruption as a damage event, and why avoiding one beats mitigating one.

DIII-D's major radius is 1.7 meters; ITER's is 6.2, and plasma stored energy scales brutally with size. Size punishes mistakes. A 200-millisecond early warning on DIII-D is a curiosity; on ITER it is an insurance policy on a machine whose construction estimates start around $25 billion. PACMAN has not run on ITER, and may never run in exactly this form, but the first integrated closed-loop demonstration on any tokamak is the prerequisite for the version that eventually does.

The strongest case against

Stated at full strength, because it is genuinely strong: plasma control is not the binding constraint on fusion energy. No reactor anywhere is sitting idle. None is waiting for a better control loop. ITER's schedule, now targeting deuterium operation in 2034 and deuterium-tritium in 2039, is driven by magnets, assembly, and tritium supply, with private timelines driven by capital and licensing, so a 20-millisecond control loop on DIII-D unlocks exactly zero kilowatt-hours of fusion electricity for the next decade.

A fairer version of the bull case: PACMAN de-risks the operations phase of machines that do not exist yet, which is infrastructure, and infrastructure is precisely what a 2030s fusion fleet needs, because the industry's graveyard is full of demonstrations that never became systems. Just do not confuse a control framework with a reactor.

Limitations

Everything reported comes from DIII-D, roughly a tenth of ITER's linear scale, and generalization to larger or differently configured devices is unproven; the authors say so themselves. Tearing-mode avoidance was demonstrated in one experiment out of five. A proof of concept, not a campaign. The press materials give no performance deltas for the RL heating control or the gyrotron optimization, and the paper is paywalled so this analysis could not verify them, while the 200-millisecond figure describes one tearing-mode case, not a general disruption predictor. Finally, the months-to-days iteration math is the authors' characterization, not a controlled measurement: read the 45x as illustrative, not certified.

What to watch

Three signals, in order of importance: first, whether the modularity claim survives contact with other machines. Watch for a PACMAN-style deployment on a second device, ideally a different size class like NSTX-U or a private-sector tokamak. Second, the paper trail: when the full Nuclear Fusion text is accessible, check the actual performance numbers behind the RL heating and gyrotron results, because press-release superlatives are not data. Third, the iteration cadence: if the Princeton team is genuinely shipping new models in days, the publication rate of follow-up experiments over the next year will show it.

For fusion investors and program managers, the actionable read is to budget for control infrastructure as a line item rather than a research project, because a control stack you can update weekly is an asset while one you update quarterly is a liability. For everyone else, watch the second-machine test: if PACMAN ports cleanly, discount every model of fusion operations that assumes control stays bespoke.

The Bottom Line

Princeton handed a real tokamak's millisecond decisions to AI. Humans stayed on goals and review. The loop runs in 20 milliseconds; the tearing-mode warning arrives 200 milliseconds early, which is ten control cycles of lookahead where a human gets zero. Twelve gyrotron knobs got jointly optimized in real time, and the second model plugged in roughly 45 times faster than the first. Not one watt of fusion power, and none this decade. But every future reactor will need this layer: control fast enough to see the instability coming, infrastructure modular enough to improve weekly. The plasma is still the hard part, but the steering is not.

Sources

Related