Measured acceleration · Physics

Deep reinforcement learning controls plasma shapes in a research tokamak

Reinforcement-learning plasma-shape control deployed on a research tokamak.

Summary

The Nature paper reports a deep reinforcement-learning architecture for magnetic control of plasmas on the TCV research tokamak. The system learned in simulation and was deployed on hardware to produce and control several plasma configurations, including elongated, negative-triangularity and snowflake shapes.

AI role

Learned control policies in simulation and transferred them to real tokamak hardware.

Narrative role

This event shows AI affecting experimental physics control, not just offline modeling or literature work.

Caveat

The result is a controlled tokamak demonstration and not a solved fusion-energy workflow.