München 2019 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

P: Fachverband Plasmaphysik

P 5: Helmholtz Graduate School I

P 5.2: Vortrag

Montag, 18. März 2019, 14:30–14:55, HS 21

Machine learning approximations of Bayesian models — •Andrea Pavone1, Jakob Svensson1, Andreas Langenberg1, Sehyun Kwak1, Udo Hoefel1, Novimir Pablant3, Matthias Brix2, and Robert C. Wolf1 for the The Wendelstein 7-X Team collaboration — 1Max-Planck-Institut für Plasmaphysik, Teilinstitut Greifswald, D-17491 Greifswald, Germany — 2Princeton Plasma Physics Laboratory, 08540 Princeton, NJ, US — 3Culham Centre for Fusion Energy, Culham Science Centre, Abingdon OX14 3D8, UK

Neural network (NN) models are trained as approximations of Bayesian models for fast data processing, opening the way to the possibility of quick inter-shot data analysis in cases where it was not possible due to computation time limitations. The NN models were tested on different diagnostic data for the inference of plasma parameters such as electron and ion temperature profiles at W7-X and JET experiments. The Bayesian models upon which they are based are developed within the Minerva framework: a common framework for modeling complex systems allows to formalize the training procedure in a way that it is mostly general and abstracted way from the single, specific diagnostic model. The training data are collected exclusively by sampling from the joint distribution of the model, so that the trained NN constitutes a surrogate of the full Bayesian models. The NN inferred plasma parameters are compared to the full Minerva Bayesian inference results. Moreover, in order to assess the reliability of the NN predictions, uncertainties of the NN output are calculated within a Bayesian framework of the NN training.

100% | Mobil-Ansicht | English Version | Kontakt/Impressum/Datenschutz
DPG-Physik > DPG-Verhandlungen > 2019 > München