AI System Controls Fusion Reactor Plasma Faster Than Any Human Operator

In some fusion systems, plasma hotter than the core of the sun can become unstable within thousandths of a second, far too fast for a human operator to respond. Researchers at the U.S. Department of Energy's Princeton Plasma Physics Laboratory (PPPL) and Princeton University have built a software framework called PACMAN (Prediction And Control using MAchiNe learning) that uses artificial intelligence to make those rapid decisions while keeping people in charge of setting the system's objectives. The work is described in a new paper in the journal Nuclear Fusion.

Fusion reactors called tokamaks use powerful magnetic fields to confine plasma, an electrically charged gas, which must stay hot, dense and stable. Small disturbances can grow within milliseconds and disrupt the reaction, and the advanced simulations used to plan experiments take days or months to run, far too slow to guide a test that may last only minutes. PACMAN instead combines several machine-learning models in a repeating control loop that runs in about 20 milliseconds, compared with the few seconds a focused human operator needs to respond.

Researchers tested PACMAN in five experiments on the DOE's DIII-D National Fusion Facility tokamak in San Diego. The system let a reinforcement-learning model take full control of the heating systems, predicted sudden bursts of energy from the plasma's edge, detected and controlled particle-driven waves in the plasma, adjusted the plasma's density and rotation to researcher-set targets, and predicted a disruptive instability called a tearing mode about 200 milliseconds before it happened, allowing the plasma to be adjusted to avoid it.

PACMAN also coordinated all six of DIII-D's gyrotrons, the microwave-based systems that heat the plasma, adjusting their power and repositioning their mirrors in real time to meet targets set in advance. "There was no algorithm to find that optimal solution before," said Andy Rothstein, a Princeton graduate student and co-lead author of the paper.