Princeton's PACMAN AI Learns to Steer Fusion Plasma Hotter Than the Sun

In some fusion energy experiments, particles reach temperatures hotter than the Sun's core, and disturbances in that plasma can develop within a few thousandths of a second — too fast for a person 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 AI to make those rapid decisions while leaving overall goals in human hands, described in a new paper in Nuclear Fusion.

Machines called tokamaks use strong magnetic fields to confine plasma — an electrically charged gas — hot and stable enough for fusion, but instabilities can grow within milliseconds. Full physics simulations can take days or months to run, far too slow to guide a live experiment. "Machine learning models can describe the plasma behavior very well, and importantly, they are the only way we have to model the plasma in millisecond times," said co-lead author Hiro Farre Kaga of the Princeton Program in Plasma Physics. PACMAN links several such models into a four-stage loop — collecting measurements, predicting plasma behaviour, deciding on actions, and applying safety-checked commands — that co-lead author Andy Rothstein said "typically runs in about 20 milliseconds," repeating continuously.

Researchers tested PACMAN in five experiments on the DIII-D National Fusion Facility tokamak in San Diego. It gave a reinforcement-learning-trained AI model full control of the heating systems, predicted sudden energy bursts from the plasma edge, identified and controlled plasma waves from fast-moving particles, adjusted plasma density and rotation to researcher-set targets, and coordinated heating from all six of the facility's gyrotrons — microwave systems that heat the plasma — at the same time.

Most notably, PACMAN predicted an instability called a tearing mode about 200 milliseconds before it would have developed, letting the plasma be adjusted to avoid it. Traditional controllers, Farre Kaga said, "can only identify this type of instability once it has already begun," after which suppressing it "can come with a lot of performance degradation."