A tokamak uses powerful magnetic fields to confine a plasma hotter than the Sun’s core, while heating beams, magnets and gas injectors continually shape its behaviour. Two-tenths of a second is a brief warning for a person, but inside this experiment it was long enough for an AI controller to redirect heating and change the plasma before a damaging instability formed.[1]
That result is one of five experiments described in PACMAN, a framework linking measurements, predictions and control. Princeton Plasma Physics Laboratory highlighted the work on 2 September following publication in Nuclear Fusion. The central ambition is to make sophisticated plasma-control models work together fast enough to influence the experiment while it is happening.[1][3]
A prediction becomes useful when something can respond
The paper’s architecture runs from diagnostic processing to actuator commands. It accommodates different approaches, including reinforcement learning, predictive control and a state-machine controller for tearing modes. This matters because fusion experiments involve several connected behaviours rather than one variable waiting to be optimised.[2]
A predictor can estimate what is about to happen; a controller chooses an intervention; a physical actuator changes the experiment. Those are different tasks inside one loop. If a warning arrives without a way to respond, it is a forecast. Couple it to an effective intervention and the trajectory of the experiment can change.
A tearing mode disrupts the magnetic structure holding the plasma. Magnetic “islands” open routes for heat to escape; if they grow and stop rotating, performance can deteriorate and a disruption can damage the machine. Microwave systems can help stabilise the plasma. PACMAN’s reported advance is to act on a warning before the mode forms, rather than waiting to suppress it after detection.[4][1]
PACMAN’s timing and control scope
| Reported quantity | Value | Role in the experiment |
|---|---|---|
| Typical PACMAN control cycle | About 20 milliseconds | Processes measurements, models and controller outputs repeatedly.Source 1Source 2 |
| Tearing-mode warning horizon | About 200 milliseconds | Reported advance warning used to intervene before tearing-mode onset; an averted event has no subsequently observed onset time.Source 1Source 2 |
| Heating system coordinated | Six gyrotrons | PACMAN retargeted mirrors and adjusted power across all six microwave heating systems.Source 1 |
| Integrated demonstrations | Five experiments | The framework was exercised across several prediction and control tasks on DIII-D.Source 1Source 2 |
One experimental platform for several kinds of intelligence
The research describes five applications: reinforcement-learning control of advanced plasma operation; prediction of edge-localised energy bursts; control of waves driven by fast particles; control of plasma profiles; and tearing-mode prediction linked to a controller. These are specialised models working on measured physical signals, not a conversational model issuing improvised instructions.[2]
A shared platform makes comparison and revision more practical. A team can change a predictor without redesigning every stage that supplies its measurements. Another can introduce a controller while using the same route to the machine’s actuators. That is an engineering advantage for a research facility where the next experiment is often intended to challenge the previous model.
PACMAN also includes a final stage that resolves competing controller requests and enforces hardware limits. In the reported experiments, human researchers set the goals and permitted operating conditions. The architecture gives specialised AI a fast path to action within those physical constraints.[1]
The gain includes the experiments you can now attempt
Fast feedback changes the questions an experimentalist can ask. Instead of preparing one fixed command sequence, they can investigate an operating strategy that adapts as the plasma evolves. They can compare how competing objectives interact, then use the recorded result to improve the next controller. That is an interpretation of the framework’s research value, not a claim of commercial fusion output.
The published paper is particularly valuable because it describes an integrated implementation, from incoming measurements to final commands. The contribution is therefore reusable experimental machinery as well as particular model results. Its examples show how learning-based and more structured controllers can coexist within the same system.[2]
A power-producing fusion plant requires many advances beyond this framework. PACMAN’s immediate contribution is practical research infrastructure: models can be added or exchanged, control requests can be reconciled and physicists can attempt experiments that need millisecond prediction and coordinated actuation. That shortens the path from an AI idea to a measurable plasma response.[1][2]
Sources & further reading
- PACMAN AI framework for controlling fusion systems safely makes key decisions in milliseconds Princeton Plasma Physics Laboratory · 2 Sep 2026 · Accessed 14 Oct 2026
- Enabling Integrated AI Control on DIII-D: A Control System Design with State-of-the-art Experiments Rothstein and colleagues / arXiv · 11 Nov 2025 · Accessed 9 Oct 2026
- Enabling integrated AI control on DIII-D: a control system design with state-of-the-art experiments Rothstein and colleagues / Nuclear Fusion · Accessed 9 Oct 2026
- The surprising attractiveness of a hurdle to developing safe, clean and carbon-free energy Princeton Plasma Physics Laboratory · Accessed 14 Oct 2026