Buffio.

A field journal of machine-made science

← All discoveries

Physics · 16 February 2022

A learned controller holds a tokamak plasma in new shapes

A deep reinforcement-learning controller, trained in simulation and deployed on the TCV tokamak, magnetically controlled a range of plasma configurations, including a two-plasma “droplet,” without further tuning on the machine.

Field
Physics
Date
16 February 2022
AI system
Deep reinforcement-learning plasma controller
Institution
DeepMind and the Swiss Plasma Center, EPFL
Published in
Nature
Identifier
10.1038/s41586-021-04301-9

A tokamak confines a hot plasma with magnetic fields so that the conditions for fusion can be approached. The shape and stability of that plasma depend on a fast, carefully coordinated set of magnetic coils — a control problem that is usually solved with carefully engineered controllers.

In a paper published in Nature on 16 February 2022, researchers at DeepMind and the Swiss Plasma Center at EPFL trained a deep reinforcement-learning agent in a plasma simulator, then deployed the learned controller directly on the TCV tokamak, at 10 kHz, without further tuning on the hardware.

The controller successfully produced a range of configurations: conventional, elongated, negative-triangularity, snowflake, and a dual-plasma “droplet.” This is a control result rather than a newly measured law of plasma physics. It is in this record because the policy that held the plasma — including configurations that had been difficult to command — was discovered by a learning system and then shown to work on a real tokamak.