Plasma Guard

In September 2026, Princeton's PACMAN AI (Prediction And Control using MAchiNe learning) ran live inside the DIII-D tokamak, completing a full sensor → predict → actuate cycle in ~20 milliseconds and forecasting plasma instabilities 200 ms before they formed. Toggle the AI off to watch instabilities grow unchecked.

Risk

0%

instability

AI Cycles

0

×20 ms

NBI Heating Power

65% · 52 MW

Plasma Density

50% · 1.0 ×10¹⁹ m⁻³

Control

● plasma · ● instability · ● quenched · -- AI beam · ◑ 20 ms cyclePrinceton PPPL · PACMAN · September 2026

Cycle time

20

ms / cycle

Prediction horizon

200

ms ahead

DIII-D experiments

5

real shots

Instabilities quenched

0

this session

PACMAN pipeline · sensors → state estimate → prediction → actuation → stable plasma

PACMAN fusion AI · cycles every 7 s

PACMAN: 20 ms per Cycle

Princeton's PACMAN (Prediction And Control using MAchiNe learning) completes a full sensor-read → state-estimate → predict → actuate loop in approximately 20 milliseconds — 50 complete control cycles every second. That is orders of magnitude faster than any human operator and even faster than most conventional software controllers, because the neural network evaluates entirely in memory rather than solving physics equations numerically.