PACMAN takes control of the heating of a tokamak and cuts an instability 200 milliseconds before it appears

Tested five times on the DIII-D tokamak in San Diego, the software framework at the Princeton Plasma Physics Laboratory reads and commands in about twenty milliseconds, whereas a human operator reacts in seconds.

PACMAN — for Prediction And Control using MAchiNe learning — is an assembly of several machine learning models designed at the Princeton Plasma Physics Laboratory, a laboratory of the U.S. Department of Energy, in collaboration with Princeton University. Its particularity is not to add another model but to wire together the complete chain: read the plasma measurements in real time, draw a prediction, and then issue the control commands. The whole thing loops in about twenty milliseconds, an order of magnitude beyond the reach of a human operator whose reaction time is measured in seconds.

The tests took place on the tokamak at the DIII-D National Fusion Facility in San Diego in five experiments. In one experiment, the system anticipated destructive instability about 200 milliseconds before it manifested itself and changed the plasma configuration to prevent it from forming. Over the entire campaign, a model trained by reinforcement learning was given complete control of the machine's heating systems.

The distribution of roles described by the authors is explicit: the decision on the fraction-second scale is entrusted to the algorithm, the definition of objectives and the safety guardrails remain in the hands of physicists. The design of the framework and these initial results are the subject of an article in the journal Nuclear Fusion.