tech
December 15, 2025
Accelerating fusion science through learned plasma control
Successfully controlling the nuclear fusion plasma in a tokamak with deep reinforcement learning
TL;DR
- A deep reinforcement learning system has been developed to autonomously control tokamak magnetic coils.
- This AI successfully contained and sculpted nuclear fusion plasma in the TCV tokamak.
- The system learned to manipulate plasma into specific shapes, aiding fusion energy research.
- Training utilized simulations due to experimental limitations, followed by validation on real hardware.
- The AI uses a single neural network to control all magnetic coils simultaneously, simplifying previous complex systems.
- Demonstrated capabilities include shaping plasma for energy generation, creating a 'snowflake' shape, and stabilizing a 'droplet' configuration with two plasmas.
- The release of the TORAX plasma simulator in May 2024 further enhances the ability to train advanced AI controllers.
- This advancement has potential applications beyond fusion, including energy efficiency and personalized medicine.
Continue reading
the original article