Deep Reinforcement Learning Achieves Breakthrough in Tokamak Plasma Control
Researchers successfully apply deep reinforcement learning to autonomously control the magnetic coils of a nuclear fusion tokamak. The new system effortlessly shapes and maintains diverse, complex plasma configurations.
Researchers introduce a novel artificial intelligence system that uses deep reinforcement learning to control the magnetic coils inside a tokamak fusion reactor. This advanced architecture autonomously learns how to shape and maintain high-temperature plasma, effectively tackling a core challenge in the pursuit of sustainable nuclear fusion energy.
The AI controller takes high-level objectives and translates them into the precise, high-frequency magnetic adjustments needed to stabilize the plasma. Unlike traditional methods, this system inherently respects physical and operational constraints while offering unprecedented flexibility in designing and testing entirely new plasma configurations without extensive manual engineering.
Scientists successfully deploy this technology on the Tokamak à Configuration Variable, where it reliably creates and manages a wide variety of plasma shapes. The system handles both standard elongated plasmas and highly advanced configurations like negative triangularity and snowflake shapes, achieving highly accurate tracking of the desired plasma targets.