Mainz 2017 – wissenschaftliches Programm
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
A: Fachverband Atomphysik
A 35: Clusters II (with MO)
A 35.4: Vortrag
Freitag, 10. März 2017, 12:00–12:15, N 3
Machine-learning assisted classification of diffraction images — •J. Zimmermann1, M. Sauppe1, A. Ulmer1, B. Langbehn1, S. Dold2, B. v. Issendorff2, I. Barke3, H. Hartmann3, K. Oldenburg3, F. Martinez3, K.H. Meiwes-Broer3, B. Erk4, C. Bomme4, B. Manschwetus4, J. Correa4, S. Düsterer4, R. Treusch4, T. Möller1, and D. Rupp1 — 1IOAP, TU Berlin — 2Univ. Freiburg — 3Univ. Rostock — 4FLASH@DESY
Short wavelength Free-Electron-Lasers (FEL) enable diffractive imaging of individual nanosized objects with a single x-ray laser shot. Due to the high repetition rate, large data sets with up to several million diffraction pattern are typically obtained in FEL particle diffraction experiments, representing a severe problem for the data analysis. We here propose a workflow scheme to drastically reduce the amount of work needed for the categorization of large data-sets of diffraction patterns, with the ultimate goal of developing an unsupervised training procedure. With a first supervised approach a classification and viewer tool is used for classifying manually selected high quality diffraction pattern. These patterns are then used as training data for a Residual Convolutional Neural Network (RCNN). The RCNN is designed for the classification of data for efficient indexing and subsequent analysis. The residual learning framework is a new type of network structure that drastically increases the depth of neural networks [He, et al. Deep Residual Learning, 2015]. First performance evaluations are done using data from a single-shot wide-angle scattering CDI experiment on silver clusters conducted in 2015 at the FLASH facility in Hamburg.