An AI controller ran a live fusion experiment and headed off an instability 200 milliseconds early

Fusion has always been a control problem, and the controls just got faster than the people watching them.

Researchers at Princeton University and the Department of Energy's Princeton Plasma Physics Laboratory plugged an AI control framework called PACMAN directly into the control system of the DIII-D National Fusion Facility. Across five experiments the software read plasma measurements and issued commands in about 20 milliseconds, where a human operator needs seconds, and it coordinated six gyrotrons at once. In one run it predicted a dangerous instability roughly 200 milliseconds before it formed and adjusted the plasma to prevent it. The framework's design and first results are published in the journal Nuclear Fusion.