AI-driven system predicts and averts fusion plasma instabilities in milliseconds
Researchers at Princeton demonstrated an AI framework that forecasted and stopped a damaging plasma instability on a tokamak within a few hundred milliseconds.
Princeton University and the Princeton Plasma Physics Laboratory created PACMAN, a modular AI framework that integrates several machine-learning models into a rapid control loop for tokamak fusion experiments. Operating on the DIII-D National Fusion Facility, PACMAN processed live sensor data, generated predictions, and issued actuator commands in roughly 20 ms, far faster than human operators. In five separate shots, the system successfully steered heating power, controlled plasma density and rotation, and, most notably, forecasted a tearing-mode instability about 200 ms in advance, preventing its onset.
Safety limits were enforced automatically, and physicists retained oversight of the control objectives. The developers say the architecture can be adapted to other tokamaks and future fusion reactors, offering a reusable AI infrastructure for the broader fusion community.
Why it matters
Real-time AI control could accelerate fusion research by preventing fast-acting plasma disruptions that hinder energy production.
In this story
