AI steers tokamak plasma in 20-millisecond control loops

Plasma inside a tokamak can change in a few thousandths of a second. A small disturbance may grow quickly enough to degrade an experiment before a human operator can react. A team from Princeton University and the Princeton Plasma Physics Laboratory has therefore connected several artificial-intelligence models directly to the control system.

The framework is called PACMAN, short for “Prediction And Control using MAchiNe learning”. It was tested in five experiments on the DIII-D tokamak. These were not isolated simulations: the models received live measurements, made predictions and sent commands to the equipment while the machine was operating.

A complete control loop in about 20 milliseconds

PACMAN first collects temperatures, densities and magnetic signals from the tokamak’s diagnostics. It checks the data and distributes it to the models that need it. Those models estimate what the plasma is doing or what it may do next. Controllers then calculate a possible action, such as changing a heating beam, gas injection or magnetic settings.

The full loop typically takes about 20 milliseconds, allowing roughly fifty decisions each second. That speed is essential. The most detailed physics simulations can require days or months of computing time, which makes them unsuitable for corrections during a live experiment.

Five experiments with five different tasks

The same architecture supported several specialised models and controllers. During the trials, PACMAN was used to:

• give a reinforcement-learning controller command of the heating systems;
• predict sudden energy bursts at the plasma edge;
• detect and control waves driven by fast particles;
• steer plasma density and rotation towards targets chosen by the researchers;
• predict a tearing-mode instability and prevent it before it formed.

In the last experiment, a model warned of the risk about 200 milliseconds in advance. That interval may sound tiny, but it gives the framework time for roughly ten complete control cycles. A conventional controller generally responds after the instability has begun, potentially at a greater cost to performance.

AI operates inside independent safety limits

The architecture keeps models, controllers and hardware output in separate layers. If two components request incompatible actions, the output layer resolves the conflict. It also enforces strict hardware limits, preventing a model from passing a command outside the permitted range.

Physicists continue to set the objectives. They choose the parameters, review the results and tune the controllers between experiments. The system’s value lies in monitoring many signals and repeating a bounded decision very quickly, rather than in unsupervised autonomy.

Reusable infrastructure for fusion research

Its modular design allows researchers to add or replace a model without rebuilding the entire control chain. That may make the framework easier to transfer to other tokamaks, although every facility has its own dimensions, sensors and actuators.

The results do not mean that a commercial fusion power plant is ready. They do show that machine-learning control can operate on a real facility within the timescale imposed by plasma physics while remaining constrained by separate technical safeguards.

The detailed results were published in Nuclear Fusion. The Princeton Plasma Physics Laboratory also describes the five experiments and the continuing role of human operators.