DeepMind RL controller shapes TCV plasmas and sustains droplets

AI system

System: DeepMind deep reinforcement learning magnetic controller for TCV (MLP policy commanding 19 coils)
Model: Not assessed
Release: Not assessed
Application developer: DeepMind, with the Swiss Plasma Center, EPFL
Role: Not assessed
Configuration: Not assessed
Inputs and outputs: Not assessed
Autonomy: Not assessed
Workflow: A high-fidelity simulator (FGE) and equilibrium reconstruction (LIUQE) of the target tokamak available for training. A flexible research tokamak with 19 independently actuated coils. A real-time control system able to run a neural network policy at the required rate.
Human oversight: Not assessed
Relation to AI: The reinforcement learning policy directly issued the coil voltage commands during the discharges, so the configurations are its output. The evidence is the peer-reviewed paper by the maker (DeepMind) and the operator (EPFL) plus an independent news report quoting an outside expert. Developer and deployer collaborated on the paper.
AI contribution: The reinforcement learning policy directly issued the coil voltage commands during the discharges, so the configurations are its output. The evidence is the peer-reviewed paper by the maker (DeepMind) and the operator (EPFL) plus an independent news report quoting an outside expert. Developer and deployer collaborated on the paper.

Participants: Developer: DeepMind, with the Swiss Plasma Center, EPFL. Operator: Swiss Plasma Center, EPFL (TCV).
Jurisdictions: Benefit: Lausanne, Switzerland. Developer: United Kingdom. Operator: Switzerland.
Ownership and control: Not assessed
Dependencies: Not assessed
Cross-border connections: Not assessed
Geopolitical implications: Not assessed

Conditions for replication: A simulator accurate enough for zero-shot transfer to the device. Operator-specified targets within the device's physical and operational constraints. Access to machine time on a research tokamak.
Replication attempts: 1 follow-up entries. Benefit recurred: 1.
Later findings: 21 July 2023. DeepMind and the Swiss Plasma Center reported upgraded RL controllers with up to 65% better shape accuracy in simulation and new validating experiments on TCV.
Availability: Not assessed
Institutional decision: Not assessed
Grounds: Not assessed
Action and result: Not assessed
Success criterion: Not assessed

Assessed benefits: Organizations. Society.
Depth: Organizations: 3. Society: Not assessed.
Reasons: Organizations: The Swiss Plasma Center gained a controller the paper calls a notable advance for tokamak feedback control. It brought less design effort for new configurations and new droplet experiments. Society: No societal outcome beyond research progress is documented.
Observed or intended: Not assessed
Rubric: Not assessed
Assessor and date: Not assessed
Alternative assessments: Not assessed
Fill: Independently reported

Kind: Not assessed
Related entries: Not assessed
Field: Energy research (magnetic confinement fusion). Tokamak plasma control.
Purpose: The controller was built to shape and maintain plasmas in TCV from high-level objectives; the configurations produced, including droplets, were its intended use.
Dates: Date basis unassessed: 2022 (precision unassessed). Updated: 21 October 2026.
Stage: Not assessed
Version: SHA-256: 067f32980f20…

Peer-reviewed study Developer

Degrave, J., Felici, F., Buchli, J. et al., 'Magnetic control of tokamak plasmas through deep reinforcement learning', Nature 602, 414–419 (2022), doi:10.1038/s41586-021-04301-9
Preprint Developer

Tracey, B. D., Michi, A., Chervonyi, Y., et al. (2023). Towards practical reinforcement learning for tokamak magnetic control. arXiv:2307.11546, 21 July 2023 (v2 5 October 2023).
News report Independent author

Will Douglas Heaven, 'DeepMind's AI can control superheated plasma inside a fusion reactor', MIT Technology Review, 16 February 2022