AI Framework Controls Fusion Plasma in Milliseconds at DIII-D

Millisecond Decisions for Million-Degree Plasma

Researchers at Princeton Plasma Physics Laboratory and Princeton University have developed an artificial intelligence framework that makes split-second adjustments to fusion reactors, operating at speeds no human could match. The system, called PACMAN (Prediction And Control using MAchiNe learning), successfully controlled plasma in five experiments at the DIII-D National Fusion Facility tokamak in San Diego.

Inside tokamaks—devices that use magnetic fields to contain plasma hotter than the sun's core—instabilities can develop in milliseconds. Traditional computer simulations that model plasma behavior take days or months to run, making them useless for real-time control during experiments that last only minutes. PACMAN solves this problem by coordinating multiple machine learning models that can analyze conditions and issue commands in approximately 20 milliseconds.

Why it matters

Fusion energy research has long struggled with the speed mismatch between plasma dynamics and human reaction times. While individual AI models have shown promise in fusion control, PACMAN represents the first modular framework that allows different models to work together seamlessly. This infrastructure approach means researchers can add, remove, or update individual components without rebuilding the entire system—potentially accelerating the path to commercially viable fusion power by enabling faster iteration on control strategies.

How the Framework Operates

PACMAN functions as a four-stage assembly line. First, it collects real-time measurements from the tokamak, including temperatures, densities, and magnetic signals. Second, it validates these readings and packages them for analysis. Third, AI models process relevant data to predict plasma behavior. Finally, controllers calculate necessary adjustments—such as increasing heating beam power—while an output stage resolves conflicts between controllers, enforces hardware safety limits, and transmits commands to the tokamak.

The modular design allows each model and controller to operate independently. Researchers can introduce new capabilities without disrupting existing functions.

Demonstrated Capabilities

During testing at DIII-D, PACMAN accomplished several firsts. It allowed a reinforcement learning model to assume complete control of heating systems. It predicted sudden energy bursts from the plasma's edge and detected waves driven by fast particles. The system adjusted plasma density and rotation to researcher-specified targets.

Most significantly, PACMAN predicted a tearing mode instability roughly 200 milliseconds before it occurred and modified plasma conditions to prevent it entirely. Conventional controllers can only respond after such instabilities begin, often causing performance degradation. The framework also simultaneously steered all six of DIII-D's gyrotrons—microwave heating systems—to achieve complex goals by retargeting mirrors and adjusting power in real time.

Built for Adaptability

The researchers emphasize that installing the first model required months of work, but adding the second took only days. This rapid deployment capability matters for a research facility where experimental conditions constantly change. Graduate student Andy Rothstein noted that the ability to retrain and redeploy models within a week enables iteration previously impossible.

Despite its autonomy, PACMAN keeps humans in control. The framework enforces safety limits regardless of AI suggestions, and physicists review each experiment to tune controllers for subsequent runs. The modular architecture is designed to work across different tokamak designs, including facilities not yet built.

These findings were first reported in the journal Nuclear Fusion by researchers at Princeton Plasma Physics Laboratory and Princeton University.