Erlangen 2018 – scientific programme

Parts | Days | Selection | Search | Updates | Downloads | Help

P: Fachverband Plasmaphysik

P 11: Helmholtz Graduate School - Poster

P 11.19: Poster

Tuesday, March 6, 2018, 16:15–18:15, Redoutensaal

Deep Learning for Plasma Diagnostics — •Francisco Matos, Florian Hendrich, Frank Jenko, and Tomas Odstrcil — Max Planck Institute for Plasma Physics, Garching, Germany

At the ASDEX-Upgrade Tokamak, a Soft X-ray diagnostic exists which can be used to perform tomographic inversion (that is, reconstructing the 2D emissivity profile) of the plasma. However, state of the art tomographic algorithms require manual tuning to detect faulty measurements, can be too slow for real-time use, and do not always produce the most accurate profiles.

Our focus is on exploring the application of Deep Learning to convert this soft-x ray data into the full images (reconstructions) of the plasma emissivity profile, with the ultimate goal of producing accurate, noise-resistant, tomographic reconstructions at real-time speeds. The current approach consists in turning the concept of a Convolutional Neural Network upside down, with fully-connected layers processing the input signal, and transpose convolutional layers responsible for learning the image features to generate. We train this network by calculating the loss as the absolute error with respect to the pixels of existing tomograms. For this, we have a dataset consisting of approximately 120 000 measurement/image pairs.