Machine learning-based preemptive RMP control for ELM-crash suppression in KSTAR plasmas

The ELM-crash suppression is a critical issue to ITER in that even a single ELM crash can severely reduce the lifetime of the plasma-facing components. The L-H transition and the onset of the ELMing occur within a somewhat predictable range but at an arbitrary moment that is difficult to know precisely. Due to such characteristics of the H-mode, the ELM-less H-mode can be achieved by automation of preemptive and event-driven control rather than feedforward approaches. Our innovative way [1] for the preemptive control is to apply RMP in the pedestal build-up period at low β_N using a real-time machine learning algorithm. As a result, we can avoid the screening of RMP from the strong ExB and electron diamagnetic flow as expected from a fully developed H-mode pedestal at high β_N. Furthermore, after applying RMP to the pedestal build-up period, the pedestal width is wider than the conventional RMP-ELM suppression discharges, while the gradient in the core region is relatively steep and high. Therefore, the shots sustain plasma performance steadily without rapid performance collapse or steady performance degradation except for the RMP effect.

[1] Giwook Shin et al., “Real-time classification of L-H transition and ELM in KSTAR”, Fusion Engineering and Design 157 (2020) 111634