SCIENCE & TECH
Somewhere inside a doughnut-shaped machine in San Diego, a cloud of gas hotter than the core of the sun is trying to tear itself apart. It has been trying for about a fifth of a second. It does not know that it has already lost.
Because for the first time in the history of fusion energy research, something was watching that could see the disaster coming — and it was not a person. It was software. It saw the instability roughly 200 milliseconds before it existed, nudged the plasma sideways, and the instability simply never happened.
That is the headline result from PACMAN, a new artificial-intelligence control framework built by researchers at the U.S. Department of Energy’s Princeton Plasma Physics Laboratory (PPPL) and Princeton University. The design and its first five real-world experiments were published this month in the journal Nuclear Fusion, and the story exploded across science media over the weekend after ScienceDaily pushed it to its front page on September 6 under a headline that did not undersell it: AI can now control fusion plasma faster than humans can react.
It is not hype. It is, if anything, understated — and it may turn out to be one of the quiet turning points in humanity’s seventy-year attempt to build a star in a box.
The problem: a gas hotter than the sun, and nobody fast enough to catch it
Fusion is the reaction that powers the sun: smash light atomic nuclei together hard enough and they fuse, releasing enormous energy. Do it on Earth, reliably, and you get essentially unlimited electricity with no carbon emissions and no long-lived radioactive waste. That is the prize, and it is the reason governments and private investors have poured billions into the field.
The most-developed approach is the tokamak — a doughnut-shaped vacuum chamber wrapped in colossal magnets. Inside it sits plasma: an electrically charged gas so hot it is usually called the fourth state of matter. The magnets hold it in place, because nothing physical could touch it and survive.
And here is the catch that has haunted fusion for decades. Plasma is not a docile fuel sitting in a tank. It is a violent, twitching, self-organising fluid that constantly tries to escape. Physicists call the escape attempts instabilities. Small ones grow into big ones in milliseconds. Big ones end the shot, dump the energy, and can damage the machine.
Keeping the plasma hot, dense and stable means constantly adjusting heating beams, magnetic fields and gas injectors — hundreds of tiny corrections, all the time. Which raises the obvious question: who, exactly, is supposed to make those corrections?
Not a human being. Andy Rothstein, a graduate student in Princeton University’s Department of Mechanical and Aerospace Engineering and co-lead author of the paper, puts the gap in brutal terms.
"A really focused human operator can respond on the order of seconds," Rothstein said. "The whole PACMAN framework typically runs in about 20 milliseconds, and it’s not running once. It’s running again and again and again. It can see small things happening in the plasma and adjust in a way that a human would never be able to do."
Twenty milliseconds against several seconds. That is not an improvement. That is a different category of reflex — closer to the difference between a person and a housefly than between a person and a faster person.
So why not just simulate it?
Because the simulations are magnificent and hopelessly slow. The sophisticated physics codes that model plasma behaviour can take days or months to run. That is excellent if you are planning next year’s experiment. It is useless if the experiment lasts a few minutes and the crisis lasts a few thousandths of a second.
This is where machine learning enters — not as a buzzword, but as the only tool that fits the clock.
"That’s great for preparing for the next experiment in a year, but for control we need models that make a decision in the moment," said co-lead author Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics, a joint programme of Princeton University and PPPL. "Machine learning models can describe the plasma behaviour very well, and importantly, they are the only way we have to model the plasma in millisecond times. The speed of these models is what’s key for control."
Meet PACMAN
The name is a stretched-out acronym — Prediction And Control using MAchiNe learning — and yes, the researchers know exactly what it sounds like. The joke is better than they let on: the thing genuinely does spend its life chasing a maze of hazards it must swallow before they multiply.
What makes PACMAN different from previous AI-in-fusion demonstrations is not that it is smarter. It is that it is plumbing.
Machine learning has been sniffing around fusion control for years, and it has scored real wins — a landmark 2024 Nature paper from a Princeton-led team, also involving Egemen Kolemen, used deep reinforcement learning to steer DIII-D plasmas away from tearing instabilities. But almost every such effort was a one-off, built from scratch, hand-wired into the machine, incapable of talking to the next model down the line. Fusion does not need one clever model. It needs a dozen models running simultaneously, watching different parts of a machine that is misbehaving in several ways at once.
"We developed this framework so that models could communicate, outputs from those models could be shared and we could do exciting physics in one integrated system," Rothstein said.
The assembly line inside the machine
PACMAN works like a four-station factory line, and it does the whole circuit roughly fifty times a second.
Station one gathers every real-time measurement the tokamak produces — temperatures, densities, magnetic signals — and bundles them together. Station two checks those readings for errors and packages them cleanly, which sounds like housekeeping but is the difference between a control system and a very fast way to break a reactor.
Station three is where the AI lives. Each model reaches into the package, grabs only the measurements it needs, and predicts what the plasma is doing — or about to do. Station four takes those predictions, turns them into commands (turn up that heating beam, move that mirror), resolves any conflicts between controllers arguing for opposite actions, enforces hard hardware safety limits, and only then sends the orders to the machine.
The final safety gate matters enormously, and the researchers put it front and centre: the framework enforces the limits no matter what an AI model recommends. The AI can propose. It cannot override the physical guardrails.
Five experiments, five different tricks
The team proved it on the DOE’s DIII-D National Fusion Facility in San Diego — the largest magnetic fusion facility in the United States and the sport’s premier test track. In five separate experiments, PACMAN:
• Let a model trained by trial-and-error reinforcement learning take complete control of the heating systems.
• Predicted sudden bursts of energy escaping the plasma’s edge.
• Detected and controlled waves in the plasma driven by fast particles.
• Steered the plasma’s density and rotation to targets set by the researchers.
• Predicted a tearing mode instability and stopped it before it formed.
That last one is the showstopper, and it deserves unpacking. A tearing mode is exactly what it sounds like: the magnetic surfaces confining the plasma break and reconnect into islands, performance collapses, and in the worst case the whole discharge disrupts. Conventional controllers cannot even see a tearing mode until it has already started.
"Then they try to suppress it, and that can come with a lot of performance degradation," Farre Kaga said. "In one of the experiments we present, a machine learning model predicts the tearing mode about 200 milliseconds in advance, so the plasma can be changed to avoid it in the first place."
Two hundred milliseconds. Roughly the time it takes you to blink. In fusion, it is a lifetime — and the difference between fighting a fire and never lighting one.
Six microwave cannons and a problem nobody had solved
There was a second result that the physicists clearly find delicious. DIII-D has six gyrotrons — systems that blast the plasma with powerful microwave beams to heat it. Aiming and powering all six optimally, at once, in real time, to hit a complex target set in advance is a genuinely hard optimisation problem. PACMAN retargeted the gyrotron mirrors and adjusted their power simultaneously, live.
"There was no algorithm to find that optimal solution before," Farre Kaga said. "When the shot ended and we looked at the data, it was doing exactly what we hoped, simultaneously moving all six in an optimal way to reach the goal."
The unglamorous result that may matter most
Ask Rothstein what surprised him and he does not name the tearing mode. He names the calendar.
Building PACMAN and installing the very first model took months of grinding work. Then they went to add the second model.
"Then we went to put in the second model, and it took a couple of days. The testing was easier, and there were far fewer bugs," he said. "DIII-D is first and foremost a research machine, and sometimes things don’t work out the way you expected. If you can put a model on in a week, you can retrain it and put a new one on the week after. It allows for iteration that wasn’t possible previously."
Months to days. That is the sound of a research field changing gear. Fusion has always been throttled not by ideas but by how slowly you can test them on machines that cost hundreds of millions of dollars and run on tightly rationed schedules. Anything that turns a year of iterations into a month of iterations is worth more than any single clever algorithm.
Expert perspective: infrastructure, not a stunt
Egemen Kolemen, associate professor of mechanical and aerospace engineering at Princeton University, jointly appointed with the Andlinger Center for Energy and the Environment and PPPL, framed the significance precisely — and the word he chose is the one that should make the fusion industry sit up.
"PACMAN uses a flexible setup where building block AI algorithms can be put together. You can add a new one, swap one out or run several at once without touching the rest of the system," Kolemen said. "That modularity is what turns AI plasma control from a series of one-off demonstrations into infrastructure the whole fusion community can build on."
Infrastructure. Not a demo, not a paper, not a press release. Roads and pipes. Because PACMAN is modular, its creators believe it can travel to tokamaks of different sizes, shapes and instrument sets — including machines that have not been designed yet. That list plausibly includes ITER, the vast international reactor under construction in southern France, and the fleet of privately funded machines that have collectively raised billions in the past few years.
The other authors on the paper are Ricardo Shousha, Keith Erickson and SangKyeun Kim of PPPL; Jalal-ud-din Butt, Peter Steiner and Azarakhsh Jalalvand of Princeton University; and Takuma Wakatsuki of Japan’s National Institutes for Quantum Science and Technology — a reminder that this is an international effort, not a single lab’s trophy. The work was funded by the DOE Office of Science and the National Science Foundation.
But who is actually in charge?
This is the question everyone asks the moment they hear "AI runs the nuclear machine," and the researchers answer it directly rather than dodging.
PACMAN does not remove people from the process. The safety limits are enforced in hardware terms regardless of what the models want. Physicists review each experiment afterwards and retune the controllers before the next one. And crucially, the goals themselves — what the plasma is supposed to do — are still set by humans.
"No matter how sophisticated your controllers, in the end it’s a human operator that sets the parameters for that control," Farre Kaga said.
It is worth being clear-eyed about what this is and is not. PACMAN is a control system, not a power plant. DIII-D is a research tokamak that does not generate net electricity, and nothing in this paper changes the enormous unsolved engineering problems standing between here and a fusion power station: tritium breeding, materials that survive years of neutron bombardment, and economics that work. What PACMAN changes is one specific bottleneck — the speed and reliability of keeping a plasma alive and well-behaved — and it changes it decisively.
What this means for you
If you are waiting for fusion electricity to show up on your utility bill, be patient: this is not that day, and anyone who tells you otherwise is selling something. But there are three reasons an ordinary reader should care.
First, the machines get better faster. Every fusion experiment that ends in a disruption is money and calendar time burned. A control system that avoids instabilities rather than mopping up after them means more useful shots per year on every machine on Earth. Compounded over a decade, that is years shaved off the schedule.
Second, this is what useful AI actually looks like. Not a chatbot writing your emails, but a stack of narrow models embedded in a control loop, running fifty times a second, with hard safety limits and a human setting the objectives. It is deeply unglamorous and it works. If you want a template for how artificial intelligence earns its place in critical infrastructure — power grids, aircraft, medical devices — this is close to the ideal shape: fast where speed matters, bounded where safety matters, and answerable to a person who decides what "good" means.
Third, it is portable. Because PACMAN is designed to be modular rather than bespoke, a control advance proven on one American tokamak in San Diego can, in principle, be lifted onto a machine in France, Japan, Britain or China. The whole field advances at once. That is rare, and it is the difference between a race and a relay.
The bottom line
For seventy years, the joke about fusion has been that it is thirty years away and always will be. The joke has aged badly in the last five, and this week it aged a little more.
A cloud of superheated matter tried to break itself in a machine in San Diego. Something noticed a fifth of a second before it started, and quietly made sure it never did. No human hand touched a control. No human eye could have seen it coming.
And then, because the framework is modular, the researchers went and plugged in the next model — in a couple of days instead of a couple of months. That second part is the one to remember.
