Plasma inside a fusion device can turn unstable within a few thousandths of a second — far too fast for a human operator to intervene. Researchers at Princeton University and the US Department of Energy's Princeton Plasma Physics Laboratory (PPPL) say they have now shown that AI can close that gap.
In a paper published in the journal Nuclear Fusion, the team describes PACMAN (Prediction And Control using MAchiNe learning), a software framework that lets several machine-learning models plug directly into a tokamak's control system and work together in a repeating loop. The whole loop typically runs in about 20 milliseconds — and it runs continuously, adjusting heating, magnets and gas injectors faster than any person could.
'Machine learning models are the only way we have to model the plasma in millisecond times,' said co-lead author Hiro Farre Kaga, a graduate student in the Princeton Program in Plasma Physics. 'The speed of these models is what's key for control.'
The framework works like an assembly line: it collects live measurements such as temperature, density and magnetic signals; checks them for errors; lets AI models estimate the plasma's current and likely future state; has controllers decide on actions such as boosting a heating beam; and finally resolves conflicting commands, applies strict hardware safety limits and sends approved instructions to the machine.
PACMAN was tested in five experiments on the DIII-D National Fusion Facility tokamak in San Diego. During those runs, an AI model trained by reinforcement learning took complete control of the heating systems; the system detected and controlled fast-particle-driven waves; it adjusted density and rotation to researcher-set targets; and it coordinated all six of DIII-D's gyrotrons at once, moving their mirrors in real time — something no previous algorithm could do.
Its most striking result involved a tearing mode, a plasma instability that can end a fusion shot. Conventional controllers only recognize a tearing mode once it has begun, then try to suppress it at the cost of performance. In one PACMAN experiment, a model predicted the instability about 200 milliseconds in advance — enough time to change the plasma so the disruption never formed.
Equally important for the field may be the framework's modularity. The first model took months to integrate; the second took days. Researchers say a new model can now be added, swapped or run in parallel in about a week, turning AI plasma control from one-off demonstrations into shared infrastructure for the fusion community. The team stresses that humans stay in charge: hardware safety limits apply no matter what a model recommends, physicists review every shot, and researchers set the objectives before each experiment.
The developers believe the framework can be adapted to other tokamaks, including machines not yet designed — one more sign that the milliseconds-long decisions inside future fusion power plants may be made by AI, with people setting the destination.
