PACMAN stops a DIII-D plasma instability with 200 milliseconds to spare
Princeton's PACMAN framework predicted a tearing mode on the DIII-D tokamak about 200 milliseconds ahead and prevented the instability by changing the plasma settings. Running in cycles of about 20 milliseconds, it handled tasks involving heating, density and plasma waves across five experiments on the same machine. The peer-reviewed results appeared in Nuclear Fusion; transfer to other tokamaks has not yet been demonstrated.
Science··Night
Control intervened before the tearing mode began
In one experiment on the DIII-D tokamak, PACMAN predicted a plasma instability called a tearing mode about 200 milliseconds before it became observable. The control system used that short interval to change the plasma settings and prevent the mode from forming. Conventional controllers detect this instability only after it starts and then try to suppress it, which can reduce plasma performance. ScienceDaily and TSM Electronique report the same early warning and preventive intervention. The work demonstrated automated control on a real fusion machine operating faster than a human operator's response, which is measured on the scale of seconds.[1], [2]
Five experiments tested different control tasks
PACMAN was used in five separate experiments on DIII-D. The framework gave a reinforcement-learning model control of the heating systems, predicted sudden bursts of energy at the plasma edge and controlled waves associated with fast particles. Other experiments steered plasma density and rotation toward targets chosen by the researchers. The system could coordinate six gyrotrons at once. Its control cycle takes about 20 milliseconds and repeats continuously, processing diagnostic data and passing model output to the machine's actuators. The study therefore tested several predictors and controllers in one infrastructure rather than only a single warning model.[1], [2]
The modular design remains untested on other tokamaks
The peer-reviewed Nuclear Fusion paper describes PACMAN as a control architecture extending from diagnostic data to final actuator commands. Predictive models and controllers operate as independent modules, allowing researchers to replace one component without disrupting the others. Safety limits remain outside the artificial-intelligence model, while people set the experimental goals. The developers say the design could be adapted to other tokamaks. All the reported experiments, however, were conducted on DIII-D, so the same speed and instability control have not yet been demonstrated on a different fusion machine.[1]
