Princeton’s PACMAN AI controlled fusion plasma in five DIII-D experiments

Researchers from Princeton University and the U.S. Department of Energy’s Princeton Plasma Physics Laboratory deployed a modular AI control framework called PACMAN on the DIII-D tokamak in San Diego, successfully controlling fusion plasma in five experiments. The framework ran reinforcement-learning, prediction, detection, and model-predictive-control systems that adjusted heating, gas injection, and electron-cyclotron-heating mirrors in real time, with one tearing-mode controller forecasting an instability about 200 milliseconds ahead and redirecting heating to avoid it. The preprint and published Nuclear Fusion paper document the architecture and experiments, while noting PACMAN's dependence on DIII-D's diagnostics and actuators, its millisecond-scale operation, and its unsuitability for sub-millisecond vertical displacement events.

Researchers from Princeton University and the U.S. Department of Energy’s Princeton Plasma Physics Laboratory deployed a modular AI control framework called PACMAN on the DIII-D tokamak in San Diego. Across five experiments, the framework ran reinforcement-learning, prediction, detection and model-predictive-control systems that adjusted heating, gas injection and electron-cyclotron-heating mirrors in real time. One tearing-mode controller forecast an instability about 200 milliseconds ahead and redirected heating to avoid it. The preprint and published Nuclear Fusion paper describe the architecture and experiments, while also documenting limits that matter for reactor deployment: PACMAN depends on DIII-D’s diagnostics and actuators, operates on millisecond-scale cycles and is unsuitable for sub-millisecond vertical displacement events.