Machine-Learning enabled analysis of ELM filament dynamics in KSTAR

Abstract

The emergence and dynamics of filamentary structures associated with edge-localized modes (ELMs) inside tokamak plasmas during high-confinement mode is regularly studied using Electron Cyclotron Emission Imaging (ECEI) diagnostic systems. Such diagnostics allow us to infer electron temperature variations, often across a poloidal cross-section. Previously, detailed analysis of these filamentary dynamics and classification of the precursors to edge-localized crashes has been done manually. We present a machine-learning-based model, capable of automatically identifying the position, spatial extend, and amplitude of ELM filaments. The model is a deep convolutional neural network that has been trained and optimized on an extensive set of manually labeled ECEI data from the KSTAR tokamak. Once trained, the model achieves a 93.7% precision and allows us to robustly identify plasma filaments in unseen ECEI data. The trained model is used to characterize ELM filament dynamics in a single H-mode plasma discharge. We identify quasi-periodic oscillations of the filaments size, total heat content, and radial velocity. The detailed dynamics of these quantities appear strongly correlated with each other and appear qualitatively different during the pre-crash and ELM crash phases.

MHD mode tracking using high-speed cameras and deep learning

“…These include but are not limit to tomographic reconstruction [24,25], equilibrium profile modeling [26][27][28][29], and the identification of reduced-order models [30][31][32]. Novel techniques such as reinforcement learning [33][34][35], reservoir computing [36], and object detection algorithms [37,38] have also been demonstrated in fusion-relevant applications.…”