Achieving ELM-suppressed Operation with the Highest Performance in DIII-D and KSTAR via Adaptive and Machine Learning Controls*

Recent experiments in DIII-D and KSTAR have achieved ELM-suppressed operation with the record performance figure of merit G=H₈₉β_N /q₉₅²>0.4 through a new integrated real-time external 3D field control. This approach integrates adaptive [1-3] and machine learning (ML) methods, automatically optimizing the ELM suppression and confinement using resonant magnetic perturbations (RMP). This approach yields the following key outcomes: 1) a up to 90% increase in G compared to the standard, non-optimized ELM suppression, 2) the establishment of nearly full ELM suppression throughout the entire discharge immediately after transitioning to H-mode, and 3) the first demonstration of fully automated 3D-field optimization through an ML algorithm. These achievements exploit the hysteresis in RMP ELM suppression access by combining an adaptive method, which iteratively adjusts the RMP amplitude to maximize confinement achievable without reverting to an ELMy pedestal, with an ML-based 3D optimizer trained on ideal plasma response computations [4], which automatically fine-tunes the RMP spectrum to maximize ELM-free phase stability. The significant increase in G is ascribed to the evolution of plasma flow observed during adaptive control. Specifically, the location of a zero-crossing layer of ExB rotation shifts radially outward to overlap with the 11/3 magnetic island, located near or at the pedestal top; this allows the maintenance of a small island and ELM suppression as RMP fields are reduced to record low values and pedestal confinement recovers. Additionally, the changes in plasma flow appear to be responsible for a reduction of turbulent transport near the pedestal top, leading to a further enhancement of the confinement. Lastly, we discuss how this phenomenon would scale in low-torque ITER scenarios. [1] F. M. Laggner et al., NF 60, 076004 [2] R. Shousha et al., POP 29, 032514 [3] S.K. Kim et al., Nature Comm 15, 3990 [4] S. M. Yang et al., Nature Comm 15, 1275