RL for Tokamak and Plasma Control
Deep RL has been applied, in published work, to control the magnetic shape of tokamak plasmas.
A hard control problem
A tokamak confines plasma with magnetic fields shaped by many poloidal field coils. Holding the plasma in a target shape and position is a fast, high-dimensional, nonlinear control problem: the plasma is unstable, the dynamics couple strongly, and control must run at kilohertz rates. Traditional control designs a separate controller for each coil and target, which is labor-intensive.
The published result
In 2022 a collaboration between a deep-learning research group and a Swiss tokamak laboratory demonstrated deep RL for magnetic control of plasma shape on the TCV tokamak. A single reinforcement-learning agent, trained in a physics simulator, directly commanded the coil voltages to achieve and hold a variety of plasma configurations, and the learned controller was then run successfully on the real device.
- A learned policy replaced a suite of hand-designed controllers
- Training happened in a free-boundary plasma-evolution simulator
- The controller shaped and stabilized real plasmas, including elongated and multi-domain shapes
Why RL fits
RL suits this problem because a reward can encode a desired shape and stability, and the agent discovers the coordinated multi-coil actuation to achieve it, including configurations that are awkward to design by hand. Learning in simulation is essential: exploration on real fusion hardware would be unsafe and slow, so this is a sim-to-real problem with a conservative deployment step, matching the safe-RL and sim-to-real principles discussed elsewhere in this library.
Relation to Kronos
Kronos machines are at design and simulation stage; the Hyperion breeder is a D-T spherical tokamak concept and the burner is a D-3He tandem-mirror concept, with construction planned to begin Q2 2027. Any learned plasma control for such devices would be developed and validated in simulation first, consistent with the published tokamak-control work and with safe-RL practice. No net-gain hardware claim is made before first-of-a-kind operation.