PACMAN AI Framework Achieves 20-Millisecond Plasma Control in Five Tokamak Experiments at PPPL

The PACMAN system integrates multiple ML models into a repeating 20-millisecond loop that ingests live tokamak signals, detects errors, estimates plasma state, and actuates heating, magnets, and gas injectors. In five separate experiments at PPPL, the framework maintained stability against instabilities that grow in milliseconds—far below human reaction times of roughly one second. This marks a shift from isolated AI controllers to a unified architecture where models share outputs in real time.

Fusion research has long faced a mismatch between slow physics simulations and the need for instantaneous decisions. PACMAN solves this by using fast surrogate models trained on prior data, enabling the system to act before instabilities disrupt confinement. The approach aligns with broader trends where AI moves from offline analysis to closed-loop operation in high-stakes physical systems such as particle accelerators and autonomous chemistry labs.

Future work must demonstrate PACMAN scaling to larger devices like ITER, where plasma volumes and actuator counts increase dramatically. Integration with physics-constrained models and formal safety verification will be required before regulatory acceptance. The current evidence rests on five short-pulse tests; longer-duration campaigns on upgraded facilities would strengthen claims of reliability.

The main limitation is the small sample of experiments and lack of reported quantitative metrics on disruption avoidance rates or energy confinement improvement.