Controlling Fusion Plasma at Machine Speed
Fusion energy holds the promise of nearly limitless power, but the process remains difficult to contain on Earth. Within a tokamak, engineers must confine plasma at temperatures exceeding the center of the Sun. This gas exists in a state where tiny disturbances grow within milliseconds, far beyond human reaction times. Researchers at the Princeton Plasma Physics Laboratory and Princeton University have built a software framework known as PACMAN to bridge this gap. This system uses machine learning to monitor and adjust fusion experiments in real time, executing decisions in roughly 20 milliseconds. The framework acts as a central nervous system for fusion machines, ensuring that hardware remains within safety limits while the AI manages complex heating and stability tasks.
Solving the Integration Problem
Previous efforts to apply machine learning to fusion suffered from a lack of standard structure. Scientists often built models from scratch for singular tasks, making it difficult to combine multiple AI agents into a single control loop. PACMAN solves this by providing a modular architecture where different models communicate and share data. The framework operates like an assembly line. It first aggregates diagnostic data including temperature and magnetic signals from the tokamak. Then, specialized models analyze this data to predict future plasma behavior. Finally, the system resolves any conflicting commands and applies safety protocols before sending instructions to the hardware. This design allows researchers to swap out components or add new models without rebuilding the entire system.
Proving the System in the Field
Testing occurred at the DIII-D National Fusion Facility in San Diego, where the team conducted five experiments. The AI demonstrated its ability to take over heating systems and adjust plasma density to target levels selected by researchers. One experiment successfully predicted a tearing mode instability 200 milliseconds before it could damage the plasma. This proactive approach marks a shift from reactive control, where traditional systems typically only act once an instability is already active. PACMAN also managed six gyrotrons simultaneously, adjusting mirror angles and power levels to achieve optimal heating outcomes that human operators could not reach on their own.
Building Infrastructure for Future Fusion
The ability to iterate quickly stands out as a primary benefit of this framework. Initially, setting up the framework required months, but adding subsequent models took only days. This speed allows for frequent updates and testing in an environment where experimental machines are often limited by time and budget. Despite the high level of automation, the researchers maintain that humans remain in charge. Operators define the parameters and set the goals for the AI, while the hardware itself enforces safety boundaries regardless of the software recommendations. The team believes this modular approach can be adapted to different tokamak designs, providing a common foundation for the wider fusion energy community to build upon as they pursue commercial-scale power generation.