Bonn 2020 – wissenschaftliches Programm
Die DPG-Frühjahrstagung in Bonn musste abgesagt werden! Lesen Sie mehr ...
Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe
T: Fachverband Teilchenphysik
T 5: Machine Learning: QCD and electromagnetic showers
T 5.4: Vortrag
Montag, 30. März 2020, 17:15–17:30, H-HS III
Understanding Generative Neural Networks for Fast Simulation of High-Granular Calorimeters — •Erik Buhmann1, Gregor Kasieczka1, Sascha Diefenbacher1, Engin Eren2, and Frank Gaede2 — 1Universität Hamburg, Institut für Experimentalphysik — 2Deutsches Elektronen-Synchrotron DESY
High-granular calorimeters are necessary for the application of particle flow algorithms in detectors for future collider projects, such as the ILD calorimeters or the CMS-HGCAL. Accurate Monte Carlo (MC) simulations of such calorimeter events demand significant computing resources. An alternative to MC is fast simulation based on generative neural networks that allow event production orders of magnitude faster than traditional MC. We are using generative adverserial network (GAN) and variational autoencoder (VAE) architectures for generating electromagnetic and hadronic calorimeter events.
Determining when the training weights converge to an optimal physics representation of the generated sampels poses a challenge when training generative models. Additionally, increasing our confidence into the accuracy of the sample generation can to be achieved by understanding the latence space represenation of physics observables. In this talk we discuss both challenges and introduce methods on how to interprete the VAE latence space in view of our physics understanding of the underlying training sample distributions.