Yanci Zhang

Magnetic shape control is fundamental to stable tokamak operation, where the plasma boundary must be maintained with high accuracy and robustness against external disturbances. Experimental observations on EAST indicate that, under large transient-error conditions, a fixed-parameter PID-based controller may drive the actuators into saturation, degrading regulation performance and potentially leading to loss of shape control. To address this limitation, this work develops a deep reinforcement learning (RL) plasma shape controller for EAST within the existing Isoflux control scheme. The controller is trained in a disturbance-rich simulation environment designed to reproduce large transient-error conditions, and employs a tailored reward design that explicitly targets large-disturbance robustness while improving early-stage learning stability under actuator constraints. Simulation results show that the learned policy achieves improved disturbance rejection and higher control accuracy in regimes where PID control tends to saturate and recover slowly. The controller is further deployed and experimentally validated on EAST, demonstrating stable plasma shape regulation with acceptable accuracy and enhanced robustness during long-pulse operation.

Recently, with the advancement of the AI field, reinforcement learning (RL) has increasingly been applied to plasma control on tokamak devices. However, possibly due to the generally high training costs of reinforcement learning based on first-principle physical models and the uncertainty in ensuring simulation results align perfectly with tokamak experiments, feedback control experiments using reinforcement learning specifically for plasma kinetic parameters on tokamaks remain scarce. To address this challenge, this work proposes a novel design scheme including the development of a low computational cost environment. This environment is derived from EAST modulation experiments data through system identification. To tackle issues of noise and actuator limitations encountered in experiments, data preprocessing methods were employed. During training, the agent collected data across multiple plasma scenarios to update its strategy, and the performance of the RL controller was fine-tuned by adjusting the weight of the integral term of the error in the reward function. The effectiveness and robustness of the proposed design were then validated in a simulated environment. Finally, the scheme was successfully implemented on EAST, effectively tracking the β _p target with lower hybrid wave (LHW) at 4.6 GHz as the actuator, and providing reference for implementing feedback control based on reinforcement learning in tokamaks.