Chinese scientists develop new AI method to control plasma in 'artificial sun'

A Chinese research team has developed a data-driven control system for tokamak plasmas that could bring sustainable fusion energy a step closer to reality, according to a paper published in Communications Physics, a journal under the Nature portfolio.

Fusion energy, often described as the "ultimate clean power," aims to replicate the process that powers the sun by fusing light atomic nuclei to release massive amounts of energy without carbon emissions. Among various reactor designs, the tokamak – a doughnut-shaped device that uses powerful magnetic fields to confine ultra-hot plasma – remains the most promising route pursued by laboratories worldwide in building an "artificial sun."

Its biggest challenge, however, is keeping the plasma stable and precisely shaped long enough for fusion reactions to produce more energy than they consume. Traditional plasma control depends on first-principles simulators and complex physics models that, while accurate, are extremely computationally demanding, making it difficult to train advanced controllers such as reinforcement-learning agents efficiently.

To address this, researchers from the Southwestern Institute of Physics, working with Zhejiang University and Zhejiang Lab, built a high-fidelity, data-driven model trained entirely on historical experimental data from the Huanliu-3 (HL-3) tokamak – China's most advanced magnetic confinement fusion device.

The model blends modern AI techniques, including one called long short-term memory (LSTM) networks – a type of recurrent neural network capable of learning long-range dependencies in sequential data – along with self-attention mechanisms and scheduled sampling. Together, these innovations allow the model to accurately predict the evolution of key plasma parameters, such as current and shape, over time, while avoiding the accumulation of errors common in traditional simulations.

The team has deployed the agent within the HL-3's real-world plasma control system. Researchers said the system remained stable and adaptable even under unfamiliar conditions, demonstrating strong robustness and "zero-shot" generalization.

The results mark a major step toward faster, more efficient training of intelligent controllers for future devices such as ITER and commercial fusion reactors, reviewers said.

Learn more about China's HL-3: China's 'artificial sun' starts new experiments with digital twin tech