Die DPG-Frühjahrstagung in Bonn musste abgesagt werden! Lesen Sie mehr ...

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

T: Fachverband Teilchenphysik

T 91: Machine Learning: Event and jet reconstruction

T 91.4: Vortrag

Freitag, 3. April 2020, 11:45–12:00, H-HS I

Di-tau mass reconstruction in ATLAS using regression-based deep neural networks — •Lena Herrmann, Philip Bechtle, Klaus Desch, Michael Hübner, and Peter Wagner — Physical Institute, University Bonn, Germany

The di-tau decay channel of resonances is important and challenging at the same time. On the one hand, it is essential for the H-analysis, but on the other hand, the unmeasured neutrinos of the tau decays complicate the mass reconstruction. As a consequence, it is hard to distinguish H-events from Z-background.

Common techniques like collinear approximations [1] or the maximum likelihood method of the Missing Mass Calculator (MMC) [2] are applied in order to estimate the invisible components and thus the invariant mass of the resonance. Alternatively, regression-based deep neural networks can be trained for this specific task. By now, the accuracy of the MMC can be approached [3] but there are still important areas of studies. Hence, edge effects, the optimum usage of the true tau-mass in the training process or the effect of tau-spin-correlations on the learning results are investigated. In the following, the optimization of a regression-based deep neural network for the di-tau mass reconstruction in ATLAS regarding the mentioned aspects, will be presented.
[1] ATLAS Collaboration: G. Aad et al., arXiv:0901.0512v4 [hep-ex]
[2] A. Elagin et al., arXiv:1012.4686, Dec 2010
[3] M. Werres, Apr 2019, Estimating the Mass of Di-Tau Systems in the ATLAS Experiment Using Neural Network Regression