Variability and risk of edge-localized modes at JET using machine learning

We present a qualitative and quantitative analysis of the behavior of edge-localized modes (ELMs) at JET. These are quasi-periodic edge instabilities in tokamaks, accompanied by the expulsion of heat and particles towards plasma-facing components. ELMs can exhibit strongly stochastic behavior, and are influenced by a combination of machine and plasma parameters. This relationship was modeled using a data-driven approach. First, more than 90000 ELMs originating from nearly 1400 JET ITER-like wall discharges were detected, by applying robust machine learning techniques. Then, the variability of ELM characteristics at stationary conditions were defined, using probability distribution of ELMs timing and size in those windows, along with descriptors of rare tail events. The latter are signs of risk, where a few occasional, but potentially impactful, large ELMs may deposit energies that well exceed the expected average ELM energy. Finally, we use regression to study how machine parameters affect those extracted features, such as the variance, percentiles and tail heaviness. The outcome is a practical overview of ELM behavior, as well as key indicators of safe machine operation.