Researchers at Princeton have built a software system that can watch a nuclear fusion experiment and issue new commands about every 20 milliseconds. The work is published in Nuclear Fusion and was tried on a real machine, the DIII-D tokamak in San Diego. A tokamak uses strong magnets to hold hot plasma.
In such machines the plasma can turn unstable in a few thousandths of a second. The software, named PACMAN, short for Prediction And Control using MAchiNe learning, is meant to handle the rapid adjustments those goals require. Machine learning can estimate what the plasma is doing in millisecond time.
PACMAN is built as a four-step loop. It first gathers live measurements such as temperature, density, and magnetic signals, checks them for errors, and packs them together. Separate models then predict what the plasma is doing or is about to do. Controllers turn those predictions into commands, for example raising a heating beam. A last stage settles conflicts among controllers, enforces hardware limits, and sends the commands to the machine. Because the pieces run on their own, a new model can be added without rewriting the rest.
Tests on a working tokamak
That design was used in five DIII-D experiments. In one, a model trained by reinforcement learning was given full control of the heating systems. In others the software predicted sudden bursts of energy at the plasma edge, detected and adjusted waves driven by fast particles, and steered density and rotation toward targets set in advance. It also predicted a tearing mode, a magnetic rip that can spoil the plasma, about 200 milliseconds before it started and changed the plasma so the rip did not form. Usual controllers only act after a tearing mode has already begun.
The same loop pointed all six of DIII-D’s gyrotrons at once. Gyrotrons are devices that heat plasma with strong microwave beams. The software moved their mirrors and changed their power in real time to meet a preset goal. After the first model took months to install, a second one was added in a few days. The authors say the same modular layout could be used on other tokamaks, including ones not yet built.