Magnetic control of tokamak plasmas through deep reinforcement learning

ABSTRACT

Nuclear fusion represents a clean, virtually limitless energy source, but sustaining the necessary plasma states inside a tokamak reactor requires complex, high-frequency magnetic control. Traditionally, this is achieved through extensive engineering and separate PID controllers for each plasma parameter. We introduce a deep reinforcement learning architecture that autonomously controls the 19 magnetic coils of the TCV tokamak to sculpt and maintain high-temperature plasmas. The RL agent, trained entirely in a physics simulator, successfully transferred zero-shot to the physical reactor, demonstrating the ability to stabilize diverse plasma configurations, including the advanced 'droplet' configuration, opening new avenues for fusion research.