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 5.6: Vortrag

Montag, 30. März 2020, 17:45–18:00, H-HS III

Towards a Data-Driven Simulation of QCD Radiation with Generative Models — André Schöning, •Christof Sauer, and Danilo Enoque Ferreira de Lima — Physikalisches Institut, Heidelberg, Germany

Recent developments in the field of machine learning open a new window on the simulation of events in high-energy particle physics through Generative Adversarial Networks (GANs) inspired by the pioneering work of Goodfellow et al in 2015. This presentation shows a potential application of GANs in terms of solely data-driven event generation with a focus on parton shower simulation. The method could be applied in analyses that are sensitive to the parton shower modelling of the background and hence rely on an accurate background estimate.
The results shown in this talk have been generated using state-of-the-art (conditional) Wasserstein GANs based on the Earth Mover’s metric. Furthermore, a comparison is made with (Gaussian) Variational Auto-Encoders (VAEs) – another avenue to generative models –, whereby the latter one shows a significantly worse performance compared to the adversarial approach. All networks presented were trained on dijet samples and W jets obtained from tt events produced with MadGraph5_aMC@NLO at LO and further processed by Pythia8.2, which serve as a surrogate to examine the applicability of the methods under well-controlled conditions. The transition to real data would be a possible next step.