Physicists at the Princeton Plasma Physics Laboratory and Princeton University have tested an artificial-intelligence system that can monitor and adjust the superheated, magnetically confined gas at the heart of a fusion reactor faster than any person could react. The framework, called PACMAN — short for Prediction And Control using MAchiNe learning — was run through five separate experiments on the DIII-D National Fusion Facility tokamak in San Diego, according to a report from the laboratory.
Each full control cycle takes the system roughly 20 milliseconds, repeating continuously through a run — fifty times a second. In one test, PACMAN predicted a "tearing mode," a plasma instability that can grow rapidly and force a shutdown, about 200 milliseconds before it would have appeared, and adjusted conditions to prevent it. Human operators typically respond on the order of seconds. The other four experiments covered reinforcement-learning control of plasma heating, prediction of edge energy bursts, and adjustments to plasma density and rotation. Findings were published in the journal Nuclear Fusion.
Why speed matters in a tokamak
Fusion reactors like DIII-D confine plasma — an ionized gas hotter than the sun's core — inside magnetic fields shaped like a torus. Small disturbances can cascade into instabilities within milliseconds, sometimes ending a run early or, in future power-generating reactors, risking damage to the vessel. Conventional controllers generally cannot flag a tearing mode until it has already begun forming, and detailed physics simulations, though accurate, take days or months to run — far too slow for real-time use. Machine-learning models trained on past plasma data approximate that behavior almost instantly, which is what makes them useful in the moment, even as physicists still rely on slower simulations for planning experiments.
Machine learning models can describe the plasma behavior very well ... they are the only way we have to model the plasma in millisecond times.
Hiro Farre Kaga, Princeton graduate student and PACMAN co-developer
Andy Rothstein, another graduate student on the project, said the system "can see small things happening and adjust in a way a human would never be able to do," according to the laboratory's account of the work. The team built PACMAN with a modular structure so new machine-learning models can be swapped in without rebuilding the whole framework — the first model took months to integrate, the researchers said, while the second took only days.
PACMAN does not operate unsupervised in any lasting sense. Physicists review the results after every run and set the objectives the system works toward beforehand. "No matter how sophisticated your controllers, in the end it's a human operator that sets the parameters for that control," Farre Kaga said. The work was funded by the Department of Energy's Office of Science and a National Science Foundation graduate fellowship, and involved researchers from Princeton, PPPL, and Japan's National Institutes for Quantum Science and Technology.
The team says PACMAN is designed to be adapted to tokamaks of different shapes and sizes beyond DIII-D, and hopes it becomes shared infrastructure other fusion labs can build on. Commercial fusion power remains years away by most estimates, and this milestone addresses plasma control rather than net energy production — but faster, more reliable control is considered a prerequisite for any reactor meant to run continuously at commercial scale.