Princeton developed an “AI guardian” that acts like a copilot for fusion reactors, predicting dangerous plasma instabilities 200 milliseconds before they strike and outpacing human operators by orders of magnitude
Inside a tokamak, fusion plasma superheated to over 100 million degrees Celsius (approximately 180 million degrees Fahrenheit) —far hotter than the core of the sun—can lose stability in a fraction of a millisecond.
When these thermal disruptions occur, magnetic control slips, and millions of amperes of electric current can slam into the reactor walls.
This razor-thin margin between harnessing clean, unlimited energy and preventing a catastrophic plasma collapse has long stood as one of nuclear engineering’s toughest hurdles.
As the Department of Energy and private industry work to commercialize clean fusion power, modular AI platforms could accelerate progress across American tokamaks.
Now, researchers at Princeton University and the U.S. Department of Energy’s Princeton Plasma Physics Laboratory (PPPL) have engineered a radical solution to bridge that gap.
A problem too fast for human hands
Managing a high-performance tokamak requires balancing dozens of volatile parameters at once. Heating systems, magnetic coils, and gas injectors must execute microscopic adjustments in real time.
Human operators simply cannot keep pace. Even the sharpest engineer requires several seconds to register a change and react—far too slow for what fusion demands.
Traditional computer simulations offer little help during live shots. While advanced models accurately depict plasma physics, they take days or weeks of compute time, rendering them useless during a short experimental run.
This fundamental speed mismatch has stymied fusion engineering for decades: the physics demands instant decision-making that neither human reflexes nor standard simulation code can provide.
Meet PACMAN: The AI framework built for fusion’s speed
To solve this operational bottleneck, the Princeton and PPPL team engineered PACMAN—short for Prediction And Control using MAchiNe learning.
Operating continuously throughout an experiment, PACMAN runs its core loop roughly every 20 milliseconds, processing diagnostic data at a frequency far beyond human capacity.
Unlike previous machine-learning efforts that addressed single, isolated plasma anomalies, PACMAN acts as a unified architecture designed specifically for multi-model orchestration.
The platform functions like a digital assembly line. It ingests stream data from thousands of tokamak sensors, checks for errors, feeds real-time predictive models, and passes optimized instructions to magnetic and heating controllers while enforcing hard hardware safety boundaries.
Five experiments, one real fusion machine
The researchers did not limit PACMAN to theoretical simulations. They deployed the framework directly onto the DIII-D National Fusion Facility in San Diego, California—the premier magnetic fusion tokamak in the United States.
Across five distinct live-plasma experiments, PACMAN demonstrated unprecedented command over critical reactor dynamics.
Demonstrations included reinforcement-learning heating control, fast-particle wave detection, and hitting precise targets for plasma density and rotation speed.
In one hallmark test, the system simultaneously coordinated all six of DIII-D’s gyrotron microwave systems, dynamically angling mirrors and tuning power output in real time.
“When the shot ended and we looked at the data, it was doing exactly what we hoped,” said co-lead author Hiro Farre Kaga, a Princeton graduate researcher. “Simultaneously moving all six in an optimal way to reach the goal.”
Predicting instability before it strikes
PACMAN’s most pivotal achievement involved anticipating tearing modes—violent plasma instabilities that disrupt magnetic field lines and trigger immediate thermal collapse.
Legacy control systems operate reactively, detecting tearing modes only after magnetic field degradation has already begun and applying heavy-handed suppression that degrades overall performance.
PACMAN transformed this dynamic entirely.
By analyzing subtle precursors in diagnostic streams, the predictive model identifies unstable conditions early enough to alter plasma parameters preemptively, sidestepping the disruption rather than fighting it.
Humans stay in charge—and iteration gets faster
Despite PACMAN’s autonomous capability, human physicists remain firmly in command. Operations teams establish target metrics before each experiment and review performance data afterward, keeping people at the center of control.
The modular framework has also accelerated the pace of U.S. fusion research. While installing the initial platform took months, integrating subsequent AI models required mere days.
Fewer integration hurdles mean faster testing and quicker refinement—a rapid development cycle previously impossible on complex research facilities.
As the Department of Energy and private industry work to commercialize clean fusion power, modular AI platforms could accelerate progress across American tokamaks.
Ultimately, by rethinking real-time control, Princeton’s “AI guardian” predicts dangerous plasma instabilities 200 milliseconds before they strike, outpacing human operators by orders of magnitude.
All the details of the study can be found here: A. Rothstein, H.J. Farre-Kaga, J. Butt, R. Shousha, K. Erickson, T. Wakatsuki, P. Steiner, S.K. Kim, A. Jalalvand, E. Kolemen. Enabling integrated AI control on DIII-D: a control system design with state-of-the-art experiments. Nuclear Fusion, 2026; 66 (7): 076050 DOI: 10.1088/1741-4326/ae7f9d
Kelly is an experienced writer with 15 years of experience exploring the big stories that shape our world, from tech breakthroughs and space exploration to climate, energy, and the fascinating quirks of science. She has a talent for turning complex ideas into sharp, memorable insights that stay with readers long after they’ve finished reading.
