SMuK 2023 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 86: ML Methods IV

T 86.3: Vortrag

Mittwoch, 22. März 2023, 18:00–18:15, HSZ/0405

Deep Neural Networks for jet-flavor tagging based on different hadronization models — •Aritra Bal, Markus Klute, and Roger Wolf — Institute for Experimental Particle Physics (ETP), Karlsruhe Institute of Technology (KIT)

Differences between the samples of either quark- or gluon-initiated jets produced by the two Monte-Carlo event generators Pythia and Herwig have been reported in the literature. A neural network can be trained to perform jet-flavor tagging on samples from either MC generator, but the performance of the network is observed to depend on the sample to which it is applied, and a network applied to a Herwig sample performs better than when applied to a Pythia sample, irrespective of the sample it was originally trained on.

We train a neural network using simple kinematic, and high-level constructed variables for better discrimination, to tag jets based on their flavor (as quark or gluon). A thorough analysis of the dependence on the input space is performed, to examine how the network responds to samples generated using different hadronization models. We also identify the critical regions of the input space where the two generators differ in the neural network response, using a Taylor Series expansion of the output function (up to 2nd order) in terms of the input variables, which we then use to find one possible answer for the generator dependence observed in the neural network application.