Nuclear fusion requires extremely hot plasma to remain stable long enough for researchers to study and control fusion reactions. Because plasma can change rapidly, human operators may not have enough time to respond to developing instabilities.
Researchers from Princeton University and the U.S. Department of Energy’s Princeton Plasma Physics Laboratory have developed an artificial-intelligence framework called PACMAN to address this challenge.
Control in Milliseconds
PACMAN combines several machine-learning models into one control system. Its control cycle can take about 20 milliseconds, allowing it to repeatedly analyse plasma conditions and adjust the system very quickly.
The framework was tested on the DIII-D National Fusion Facility tokamak in San Diego.
Predicting Instability
One notable experiment involved a plasma instability called a tearing mode. Researchers found that an AI model could predict the instability about 200 milliseconds before it developed.
That advance warning could allow the system to change plasma conditions before the instability becomes a larger problem.
AI Beyond One Task
The experiments also demonstrated AI control of heating systems, plasma density, rotation and waves produced by fast particles. The framework’s modular design allows researchers to add or replace AI models more easily.
Humans Still Set the Rules
PACMAN is not intended to remove people from fusion research. Hardware safety limits remain in place, while scientists define objectives and evaluate the experimental results.
If the approach proves reliable across different machines, AI could become an important tool in the effort to make fusion experiments more controllable and, eventually, help advance practical fusion energy.