Researchers at the Princeton Plasma Physics Laboratory (PPPL) and Princeton University have released PACMAN, a new AI framework designed to control fusion systems in real time. The system, detailed in a recent paper in the journal Nuclear Fusion, successfully executed five experiments on the DIII-D National Fusion Facility in San Diego. Unlike traditional control loops that struggle with the speed of plasma instabilities, PACMAN operates in approximately 20 milliseconds, allowing it to react to disturbances far faster than any human operator could manage.

Architecture and Modularity

PACMAN functions as a modular assembly line with four distinct stages: data gathering, error checking and packaging, AI prediction, and controller command resolution. This architecture allows researchers to plug in different machine learning models without disturbing the rest of the system. Co-lead author Hiro Farre Kaga noted that while traditional simulations take days or months, PACMAN’s ML models are the only way to model plasma behavior in milliseconds, which is critical for active control.

Real-World Performance on DIII-D

During testing at the DIII-D tokamak, the framework demonstrated significant capabilities, including predicting tearing modes 200 milliseconds before they occurred. This predictive ability allowed the system to adjust plasma conditions to avoid instabilities entirely, rather than trying to suppress them after the fact. The framework also simultaneously controlled all six of DIII-D’s gyrotrons, adjusting mirror targets and power levels in real time to achieve complex heating goals without a pre-existing optimal algorithm.

The Builder’s Advantage

For infrastructure teams, the key takeaway is the drastic reduction in iteration time. Co-lead author Andy Rothstein reported that while building the initial framework took months, adding subsequent models took only days. This modularity turns AI plasma control from one-off demonstrations into reusable infrastructure. The system enforces strict hardware safety limits, ensuring that human operators retain control over parameters while the AI handles the high-speed decision-making.

Key Takeaways

  • PACMAN runs control loops in ~20ms, significantly outperforming human reaction times.
  • The framework uses a modular four-stage pipeline to integrate various ML models.
  • Testing on DIII-D showed the system could predict tearing modes 200ms in advance.
  • Adding new models to the system reduced development time from months to days.

The Bottom Line

This is what happens when you treat fusion control as a software architecture problem. By decoupling models from the hardware loop, PPPL has built a platform that lets physicists iterate on AI strategies as quickly as they can write code, rather than waiting for months-long simulation cycles.