Zero-shot AI controller unlocks flexible, adaptive plasma control for next-generation fusion reactors

Controlling plasma shape in tokamak fusion reactors has long relied on systems which are proven yet inflexible, requiring constant manual tuning and struggling to adapt when plasma suddenly shifts. As fusion energy nears commercial viability, researchers turn to reinforcement learning (RL) to create smarter, more adaptive control systems. However, conventional RL models are task specific and require complete retraining when operational targets change.

Researchers working with China’s HL-3 tokamak have developed a breakthrough zero-shot RL controller, capable of handling entirely new plasma shapes and current profiles without additional training. Its flexibility comes from learning from experience instead of starting from scratch each time.

The system was trained on 1,115 plasma discharges from HL-3 collected between 2023 and 2025, originally controlled by PID systems. This operational data taught the AI a generalized policy understanding plasma transition physics. The approach combines Generative Adversarial Imitation Learning with a Hilbert space representation, a mathematical framework mapping plasma states by the difficulty of transitioning between configurations.

This Hilbert latent space acts as a roadmap, enabling goal-directed control through geometrical rewards reflecting time-optimal transitions. It achieves stable, precise shaping across a broad range of operating scenarios, covering shots #5000 to #11908.

The architecture includes an actor-critic model, a discriminator encouraging mimicry of PID control behaviour, and a Hilbert encoder mapping plasma states for navigation. Training in a data-driven dynamics simulator showed stable convergence and impressive trajectory tracking.

Stay ahead in the fusion revolution explore more breakthroughs from leading innovators in clean energy technology.