Ziyu GUO, Search for the Higgs boson in the tt̄H (H → bb̄) channel in the ATLAS experiment at the LHC using machine learning methods and synchronization of the ITk geometry description for simulation and radiation studies for the HL-LHC ATLAS upgrade

mardi 5 nov. 2019, 14:00 → 17:00 Europe/Paris

Amphi (CPPM)

 Résumé :
According to the Standard Model (SM), elementary particles obtain their mass by coupling with the Higgs field. This mechanism predicts the existence of the Higgs boson, which was discovered by the ATLAS and CMS collaborations in July 2012. The couplings between the Higgs and vector bosons are completely established, while the measurements of the strength of the Higgs-fermion couplings, the so-called Yukawa couplings, are incomplete. Being much stronger than the ones for the other quarks and close to unity, the Higgs-top quark Yukawa coupling is the most interesting one. It can be a hint for new physics if any measurement shows deviation from the theory. The associated production of a Higgs boson with a top quark pair (tt̄H) allows for a direct measurement of the top Yukawa coupling.

This thesis presents a search for tt̄H events with the Higgs boson decaying into a bottom quark pair (H → bb̄) based on data collected with the ATLAS detector at the LHC during Run 2 operation, with proton pair collisions at a center of mass energy of 13 TeV. The multiple jets and b-jets in the final state make the analysis challenging and heavily relying on advanced analysis techniques. The large modelling uncertainties of the tt̄ +jets backgrounds are a driving factor of the sensitivity.

In the analysis round using datasets collected in years 2015 and 2016 for an integrated luminosity of 36.1 fb−1, the author studies and optimizes boosted decision trees (BDT) to firstly solve the jet-parton assignment in the reconstruction of the tt̄H signal, and, in a second step, separate the signal from the tt̄ + jets backgrounds. The classification BDT output is one of the discriminants to fit experimental data with simulated samples based on the SM expectation. Through this process, the existence of the tt̄H signal is tested. Under the background only hypothesis, the observed (expected) significance of the tt̄H signal over the expected SM background is 1.4 (1.6) standard deviations, which is consistent with both SM background-only and tt̄H prediction.

Targeting a contribution to the new analysis round using the full Run 2 data, deep learning techniques are also explored. Recurrent neural networks are exploited, aiming to, in one step, better solve the event reconstruction and classification. This method still uses as input the physics motivated features manually calculated from the basic kinematics of the final state objects. To use these raw features as input directly, a deep neural network model is designed while incorporating the physics domain knowledge to automatically extract more discriminating features. To improve the signal sensitivity while constraining backgrounds, an event categorization, having separately the tt̄H and different tt̄ components enriched regions, is implemented using a deep neural network based multi-classifier, which is compared with the manual splitting in the previous analysis. Finally adversarial networks are studied in order to decrease the tt̄ modelling uncertainty.

The High-Luminosity Large Hadron Collider (HL-LHC), the upgrade of the current LHC, with an enhanced luminosity by a factor of 10 beyond the LHC’s design value, is expected to start operation in the mid-2020’s with increased potential for physics discovery. To cope with this new phase, the ATLAS detector will be upgraded, including a new inner tracker (ITk) which geometry is under design. For the detector simulation and radiation study, the geometry descriptions are independently implemented. The author contributes to the synchronization of the two geometries, which is important to validate a proper radiation estimation.

Membres du jury :
- co-directeur : Cristinel Diaconu, CPPM
- co-directeur : Thierry Artieres, LIS & Ecole Centrale Marseille
- co-directeur : Yann Coadou, CPPM
- rapporteure : Anne-Catherine Le Bihan, Institut pluridisciplinaire Hubert Curien, Strasbourg
- rapporteur : David Rousseau, Laboratoire de l'accelerateur lineaire, Orsay
- examinateur : Elisa Fromont, Université Rennes 1
- examinateur : Mossadek Talby, CPPM

CERN-THESIS-2019-222