Deep-learning enabled clustering of KSTAR ECEI data

A key enabler to scientific studies is the easy access to relevant measurements. As the corpus of measurement data increases geometrically, due to more measurements being taken at increasing fidelity, better ways to find relevant data are required. Approaching this problem, we explore how unsupervised machine learning models can be used to classify a corpus of electron cyclotron emission imaging (ECEI) data from the KSTAR tokamak. A dataset consisting of prominent MHD phenomena such as magnetic islands and ELM filaments is compiled and clustered using end-to-end trainable deep learning architectures. In particular we are using an approach based on generative adversarial networks and deep divergence-based clustering. On this reference dataset, the used methods achieve cluster accuracies of up to 90%.