Hannover 2020 – wissenschaftliches Programm
Die DPG-Frühjahrstagung in Hannover musste abgesagt werden! Lesen Sie mehr ...
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
K: Fachverband Kurzzeit- und angewandte Laserphysik
K 4: Poster
K 4.13: Poster
Dienstag, 10. März 2020, 16:30–18:30, f428
Quantitative coherent diffraction imaging via deep learning: from simulation to reconstruction — •Alessandro Colombo1, Julian Zimmermann2, and Daniela Rupp1 — 1ETH Zürich, 8093 Zürich, Switzerland — 2Max-Born-Institut, 12489 Berlin, Germany
Coherent Diffraction Imaging (CDI) is a lens-less technique that exploits the 2D measured diffraction pattern I(k→), produced by short-wavelength coherent radiation illuminating e.g. an individual nanoscale structure, to retrieve its electron density distribution ρ(x→). The use of Deep Learning for classifying diffraction patterns already proved to outperform standard approaches [1]. Here we present an approach based on Deep Convolutional Neural Networks for regression, to directly retrieve the electron density structure ρ(x→) when classical reconstruction algorithms do not apply, as, for example, in the wide-angle scattering regime. The generation of a sufficently large training dataset for the supervised learning of the network is done by simulating diffraction data. A model representation of the particle shapes, that enables parametrization, has to be found, with the goal of retrieving the optimized parameters from the neural network for each CDI pattern. The simulation software and the neural network are described, along with a first comparison with experimental data, showing that such a simulation-reconstruction scheme may provide a quick and quantitative insight into large CDI datasets.
[1] Zimmermann et al. Physical Review E 99.6 (2019): 063309.