Deep learning-enabled real-time prediction of impurity-induced detachment in EAST

Impurity seeding has been consistently demonstrated to facilitate plasma detachment, effectively reducing the amount of heat and particles reaching divertor targets. However, achieving and maintaining a stable detached state requires precise, real-time monitoring of the seeding rate. Current limitations in diagnostic accuracy and reliance on manual adjustments hinder this process. Here, a novel approach based on deep learning is proposed to assist in monitoring the state of detachment in the Experimental Advanced Superconducting Tokamak. This method enables instantaneous prediction of the plasma electron temperature near strike points on divertors. The model circumvents the conventional dependence on Langmuir probes for detachment control, the reliability of which will become increasingly challenging to ensure in future reactor environments. Instead, radiation data detected by photodiodes are primarily adopted to accommodate diverse operational conditions. Rigorous analysis confirms that the key determinants of the detachment state include the neutral beam injection (NBI) power, plasma current, line-averaged density, and impurity seeding rate. NBI synergizes with radio-frequency heating, broadening heat flux profiles and thereby facilitating plasma detachment. The effect of impurity seeding is consistent across different toroidal seeding locations. Despite being trained on nitrogen-seeding experimental data, the model demonstrates self-consistency with the aforementioned findings when applied to neon-seeding and argon-seeding discharges. This consistency further validates the applicability of the model across different impurity seeding scenarios. This fresh perspective will advance the understanding of detachment control.