PACMAN AI Runs the Fusion Reactor

An AI Took the Controls at 20 Milliseconds Per Cycle

Inside a fusion reactor, gas heated to ten times the temperature of the sun can become unstable in a few thousandths of a second. A human operator typically needs several seconds to notice a problem and respond. That gap has always been one of fusion energy's hardest engineering puzzles: the physics moves faster than people can. In July 2026, a team at Princeton University and the U.S. Department of Energy's Princeton Plasma Physics Laboratory published a paper in Nuclear Fusion describing PACMAN (Prediction And Control using MAchiNe learning). In five live experiments on DIII-D, the largest magnetic fusion facility in the United States, PACMAN ran the reactor's controls autonomously, cycling through sense, predict, and act in about 20 milliseconds. That is roughly 150 times faster than a human operator.

The Instability That Could Not Wait

The specific challenge PACMAN was tested against is called a tearing mode: a magnetic instability that, if it propagates, can cause the plasma to collapse entirely and dump energy into the reactor wall. Conventional systems cannot see a tearing mode until it has already begun. PACMAN predicted one approximately 200 milliseconds before it formed and adjusted the plasma proactively, preventing the instability from appearing at all. That is the difference between steering and braking. A human operator brakes when they see a hazard. PACMAN steers around it before the hazard exists. The prior state of the art required physicists to run simulations for days or months to prepare a controller, then hand it to operators for runs lasting minutes. Any instability outside the simulation's assumptions was too fast to catch.

Credit: Princeton University / PPPL.

How the Assembly Line Works

PACMAN is structured as a modular pipeline. Live measurements from the reactor's temperature, density, and magnetic field sensors flow in continuously. Machine-learning models predict what the plasma is about to do. A controller layer converts those predictions into hardware commands: raise a microwave beam's power, reposition a mirror, adjust density or rotation. A final stage checks every proposed command against hard safety limits before anything reaches the machine. In the five DIII-D experiments, PACMAN simultaneously steered all six of the reactor's gyrotrons, the microwave systems that heat the plasma. Coordinating six heating systems in real time while predicting instabilities is a task no human operator could run at 20-millisecond intervals.

Machine learning models can describe the plasma behaviour very well, said Hiro \

Farre Kaga, a graduate student in Princeton's Program in Plasma Physics. They are the only way to model the plasma in millisecond times. The remark captures why the move from conventional control to AI is not merely a speed improvement: it is access to physics that was simply uncontrollable before.

From Demo to Infrastructure

Previous AI plasma-control results were one-offs: a model trained for one reactor, one condition, one experiment. Installing a new model meant weeks of work and a restart of the control system. PACMAN is designed differently. Its modular structure means a new model can be added in days without touching the rest of the framework. After the first model took months to install, the second took a couple of days. Egemen Kolemen, associate professor of mechanical and aerospace engineering at Princeton and principal investigator, put it directly: "That modularity turns AI plasma control from demonstrations into infrastructure." The paper's authors describe plans to adapt PACMAN for tokamaks with different shapes and sizes, including machines not yet built. ITER, the international reactor under construction in France, is the obvious next-scale test.

Humans Still Set the Targets

PACMAN operates autonomously inside bounds that physicists define before each experiment. The system cannot override a safety limit, and it does not choose the plasma's goals: that remains the scientists' job. Every decision the AI makes is logged so operators can examine the run afterward and refine the next one. Before PACMAN, a control adjustment that underperformed might take weeks to analyse and rebuild. Now the team can update a model between experimental sessions and test it the following week. DIII-D runs roughly 15,000 plasma shots per year. At that volume, the gap between what physicists know in simulation and what the reactor can actually hold is, for the first time, closing faster than the plasma can escape.