Google’s AI, DeepMind Controls Plasma Shapes For Nuclear Fusion

A subsidiary of Alphabet, DeepMind, has developed artificial intelligence (AI) that has shown promise in controlling magnetic devices that keep super-hot, unruly plasma in a stable configuration, which can be used to make electricity.

Tokamaks are plasma-condensing devices made out of a succession of high-power magnetic coils that produce a vessel in which plasma may be controlled at temperatures as high as the sun’s core. Hydrogen atoms can fuse together in the presence of properly contained plasma, which is being looked at as a long-term source of power.

In a new publication published in Nature, DeepMind and EPFL’s Swiss Plasma Center (SPC) describe how they collaborated to develop a set of DeepMind AI algorithms capable of manipulating the structure of plasma inside the vessel.

For testing plasma confinement for nuclear fusion, the SPC has a “variable-configuration tokamak” (TCV) vacuum vessel. SPC needed a mechanism to make sure it could pick the proper values for each variable in the control system’s plasma confinement, such as voltage.

SPC already had a well-informed simulator, but according to SPC, its users still have to perform lengthy computations to establish the correct value for each variable in the control system.

With the correct parameters, the control system may confine plasma in a way that prevents it from colliding with the TCV vacuum vessel’s walls, which would result in plasma degradation.

The researchers claim to have developed a “before undescribed architecture for tokamak magnetic controller design that autonomously learns to command the whole set of control coils,” according to the researchers. This architecture can cut down on the time it takes to create new plasma configurations.

According to SPC’s blogpost, DeepMind’s AI can build and sustain particular plasma configurations, including “elongated, conventional shapes, as well as advanced configurations, such as negative triangularity and’snowflake’ configurations,” as outlined in the study.

The algorithms were successfully run on SPC’s genuine TVC, not only its emulator, by DeepMind and SPC. Inside the vessel, they also exhibited a sustained configuration of two distinct plasma “droplets.” DeepMind has yet to publish its research post, but it will do so soon at this link.