Princeton researchers have built an AI system called PACMAN that can monitor and control fusion plasma faster than any human operator, running a full predict-and-adjust cycle in about 20 milliseconds. In tests at the DIII-D National Fusion Facility in San Diego, the framework successfully forecast a damaging "tearing mode" instability roughly 200 milliseconds before it would have formed and adjusted the plasma to stop it.
Fusion plasmas are notoriously unstable: small disturbances can grow within milliseconds and disrupt or even end a reaction, far faster than a human can react. PACMAN (Prediction And Control using MAchiNe learning) continuously reads real-time measurements — temperature, density, magnetic signals — and uses machine-learning models to predict what the plasma will do next, then hands adjustments to controllers. "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 graduate student Hiro Farre Kaga, one of the researchers on the project.
The system ran successfully across five separate experiments, with hardware safety limits and post-experiment human review keeping operators in the loop. "No matter how sophisticated your controllers, in the end it's a human operator that sets the parameters," Farre Kaga noted. The results, published in the journal Nuclear Fusion, mark a step toward the kind of split-second automated control that commercial fusion power plants will likely need to run safely and continuously on the grid.