Researchers at the US Department of Energy's Princeton Plasma Physics Laboratory (PPPL) and Princeton University have developed an artificial intelligence-based software framework capable of making rapid decisions to control fusion plasma at speeds beyond human response.

The framework, called PACMAN, or prediction and control using machine learning, was successfully tested in five experiments at the DIII-D National Fusion Facility tokamak in San Diego, according to the researchers.

Their findings were published in the journal "Nuclear Fusion".

Fusion research seeks to replicate the process that powers the sun to produce a potentially abundant source of electricity. Tokamaks use powerful magnetic fields to confine extremely hot plasma, but keeping the plasma stable requires continuous adjustments to heating systems, magnets, and gas injectors.

Plasma instabilities can develop within milliseconds, making them too fast for human operators to address. Computer simulations can predict plasma behaviour, but some simulations take days or months and therefore cannot provide real-time control during experiments.

Machine-learning models, however, can predict plasma behaviour within milliseconds, researchers said.

PACMAN integrates multiple AI models into a single control system. It collects real-time measurements, such as plasma temperature, density and magnetic signals, checks the data and uses AI models to assess the plasma's current and predicted behaviour.

Controllers then use those predictions to determine necessary adjustments, such as changing heating power. Before commands are sent to the tokamak, PACMAN checks them against hardware safety limits and resolves conflicting instructions.

During five experiments at DIII-D, the system demonstrated several capabilities. It allowed a reinforcement-learning model to control plasma heating, predicted bursts of energy from the plasma edge, detected and controlled waves caused by fast particles, adjusted plasma density and rotation, and predicted a potentially disruptive instability known as a tearing mode.

In one experiment, the AI predicted the tearing mode about 200 milliseconds before it was expected to occur, allowing researchers to adjust the plasma to prevent the instability rather than suppressing it after it began.

PACMAN also coordinated all six of DIII-D's gyrotrons, which heat plasma using powerful microwave beams. The system simultaneously adjusted their power and mirror positions to achieve targets set by researchers.

Researchers said the framework could significantly speed up the development and testing of new AI models. While integrating the first model took months, adding a second model took only a few days, allowing researchers to test, retrain, and replace models more quickly.

The researchers stressed that PACMAN is not designed to eliminate human oversight. Human operators continue to set the control objectives and parameters, while the framework enforces hardware safety limits regardless of AI recommendations.

Egemen Kolemen, an associate professor at Princeton University and PPPL, said PACMAN's modular design could allow AI algorithms to be added, replaced, or operated simultaneously without requiring changes to the rest of the system.

The researchers believe the framework could eventually be adapted for tokamaks of different designs and sizes, potentially providing a common platform for AI-based plasma control across future fusion facilities.