Abstract: Plasma fusion is a promising source of future energy. We apply machine learning techniques to detect blob features in plasma fusion data to help physicists gain deep insight into their simulations and experiments in Tokamak reactors. The blobs, which are not mathematically well-defined features, typically form on outer regions of the reactor and can possibly damage the reactor. We thus prototyped machine learning approaches to automatically detect blobs in XGC fusion simulation data.

In this talk, I will first explain the user interface we designed for domain scientists to label data for training purposes. I will then present our data preprocessing with topology analysis and the machine learning approaches based on the UNet architecture. Preliminary results are obtained with arbitrarily labeled training data. Finally, I will outline future direction with the applied concepts.

Bio: Martin Imre is a research aide in the Mathematics and Computer Science division. He received his B.Sc. (2014) and M.Sc. (2016) in software engineering at the Vienna University of Technology. His research interests are in scientific and information visualization and high-performance computing.