Machine Learning enabled detection of ELM-Precursors in KSTAR ECEI data.

The emergence and dynamics of edge-localized filamentary structures inside tokamaks during high-confinement mode can be studied using an ECEI (Electron Cyclotron Emission Imaging) diagnostic system. This diagnostic samples a temperature map of a 2D poloidal cross-section of the plasma on a microsecond time scale. Previously, detailed analysis of these filamentary dynamics and classification of the precursors to edge-localized crashes has been done manually. We present an algorithmic approach capable of automatically identifying the position and dimension of these filaments. This is achieved using a Single-Shot Detection convolutional neural network (SSD), a type of machine learning algorithm that detects features within images and proposes specific regions in which it expects a certain object is present. The machine learning model has been trained and optimized on an extensive set of manually labeled ECEI data from KSTAR (Korean Superconducting Tokamak Advanced Research), allowing the model to "learn" the features that constitute a precursory filament. In order to be even more hands-free, the algorithm is also capable of filtering raw ECEI data and masking out bad channels automatically; reasonably inferring masked data based upon neighboring channels. This model will allow for a more efficient and broad study of edge-localized filamentary dynamics with the aim of achieving a better understanding of edge localized mode crashes.