Researchers at the U.S. Department of Energy's Princeton Plasma Physics Laboratory and Princeton University have developed an artificial intelligence (AI) software framework that can complete a prediction and control of fusion plasma in about 20 milliseconds, rapidly adjusting its state while being subject to strict hardware safety constraints. The related results were published in the latest issue of the journal Nuclear Fusion.
Application of the PACMAN AI framework in a fusion system (artist's concept). Image credit: Princeton Plasma Physics Laboratory, U.S. Department of Energy
Scientists are using devices such as tokamaks to explore controlled nuclear fusion. A tokamak confines high-temperature plasma within its interior through strong magnetic fields. For fusion reactions to continue, the plasma needs to remain at high temperature, high density, and in a stable state. However, some instabilities in the plasma develop very quickly and can intensify rapidly within a few milliseconds, often leaving human operators no time to react.
Predicting what will happen next in the plasma is key to controlling fusion reactions. Traditional computer simulations may take days or even months to run once, making them unable to meet the requirements of real-time control. Machine learning models, by contrast, can use experimental data to quickly determine the plasma state and predict its changes.
In this case, the system developed by the researchers is called PACMAN, which stands for "Predict and Control using Machine learning." It can call multiple AI models simultaneously and adjust related parameters such as heating equipment and magnetic fields based on prediction results. The system also checks the control plans given by the AI and sets hardware safety limits to ensure that the equipment does not perform operations beyond the safe range.
Results from five experiments conducted on the tokamak at the U.S. Department of Energy's DIII-D National Fusion Facility show that PACMAN can use AI to control the plasma heating system, predict energy bursts at the plasma edge, detect and control plasma waves driven by high-speed particles, and adjust plasma density and rotation states according to preset targets.
The researchers also used PACMAN to simultaneously control six gyrotrons in the DIII-D device. Gyrotrons deliver energy to the plasma through powerful microwave beams. In the experiments, the AI was able to simultaneously adjust the power and mirror positions of the six gyrotrons, enabling these devices to work in coordination to achieve the targets set by the researchers.
The researchers emphasized that AI has not replaced human control over fusion experiments. Experimental targets and control parameters are still set by scientists, and the system always enforces hardware safety limits.