Princeton's AI system controls fusion plasma in milliseconds

Summary

Researchers at Princeton and PPPL introduced PACMAN, a modular AI framework. The system monitors fusion plasma and adjusts it within approximately 20 milliseconds. In five experiments conducted on the DIII-D tokamak, PACMAN predicted tearing mode instability approximately 200 milliseconds before its onset.

Why it matters

Instabilities in fusion plasma grow on a millisecond scale. This speed is far beyond human reflexes. Therefore, real-time AI control is a prerequisite for stable and long-duration operation. PACMAN's modular design allows new models to be deployed in days rather than months. This feature could accelerate the iterative cycle that tokamak research currently struggles with. Consequently, the path to net-energy-positive machines could be shortened.

Highlights

• PACMAN completed the sense-predict-act loop in approximately 20 milliseconds. The system coordinated six gyrotrons on DIII-D simultaneously. No algorithm had previously been developed for this task.
• A reinforcement learning model was granted full authority over heating systems. A separate model predicted the tearing mode approximately 200 milliseconds in advance. This enabled proactive suppression instead of reactive correction.
• Integrating the first model took months. The second model was added in just a few days. This demonstrates that PACMAN will enable weekly retraining cycles in research tokamaks.

Critical take

All results were obtained from a single medium-sized tokamak called DIII-D and only five shots. Therefore, it has not yet been proven whether the findings can be generalized to larger and more complex devices like ITER or SPARC. The phrase 'faster than humans' also exaggerates the situation. Since human operators were not expected to close a 20-millisecond control loop.