SMuK 2023 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
T: Fachverband Teilchenphysik
T 103: AI Topical Day – Simulation, Inverse Problems and Algorithmic Development (joint session AKPIK/T)
T 103.6: Vortrag
Donnerstag, 23. März 2023, 17:00–17:15, HSZ/0004
Reconstruction of SAXS Data using Invertible Neural Networks — •Erik Thiessenhusen1, Melanie Rödel1, Thomas Kluge1, Michael Bussmann2, Thomas Cowan1, and Nico Hoffmann1 — 1HZDR, FWKT, Dresden, Germany — 2CASUS, Görlitz, Germany
The understanding of laser-solid interactions is important to the development of future laser-driven particle and photon sources, e.g., for tumor therapy, astrophysics or fusion. Currently, these interactions can only be modeled by simulations which need verification within the scope of pump-probe experiments. This experimental setup allows us to study the laser-plasma interaction that occurs when an ultrahigh-intensity laser hits a solid density target. We employ Small-Angle X-Ray Scattering (SAXS) to image the nanometer-scale spatial- and femtosecond temporal resolution of the laser-plasma interactions. However, the analysis of the SAXS pattern is an ill-posed inverse problem meaning that multiple configurations of our target might explain the same measurement due to the loss of the phase information. We approach the ambiguities of the inverse problem by a conditional Invertible Neural Network (cINN) that is returning a probability density distribution over target parameters explaining a single SAXS pattern. We will show that the domain gap between generated training and experimental data can be approached by integrating perturbations of experimental data into the training workflow. We assess the applicability of our approach to a selected set of grating targets in terms of a comprehensive evaluation on simulation and experimental data.