Pedestal optimization right after L-H transition by machine learning-based preemptive RMP application algorithm for avoiding ELM-crash event in KSTAR

The unmitigated large ELM crashes can cause severe damage to plasma-facing components (PFCs) in ITER tokamak. Especially the first large ELM crash right after L- to H-mode transition is a concern in that even a single ELM crash can also reduce the lifetime of the PFCs. In order to solve the issue, the RMP is preemptively applied to the plasma right after the H-mode transition by a machine learning-based RMP control algorithm in KSTAR [1]. In the previous study, we applied the preemptive RMP after the moment that the plasma is less affected by the shaping effect for focusing on the first ELM crash suppression. However, since we want to apply the preemptive RMP at a typical onset of H-mode transition where the plasma shape rapidly changes from circular to diverted, the ML-based method encounters a harsher condition to suppress the first ELM crash. Through the optimization process to overcome the condition, we found the optimized slew time and amplitude of the preemptive RMP. As a result, the plasma obtained after the optimization shows different plasma profile patterns from those shown in conventional RMP-ELM crash suppression experiments. In addition, after the pedestal in which the first ELM crash is suppressed is formed, the ELM suppression is maintained over 37 seconds.