Search for additional heavy Higgs bosons decaying to tau lepton pair in the CMS experiment at LHC with machine learning techniques

Despite decades of correct predictions, physicists are convinced that the Standard Model (SM) does not show us the whole picture. Among the various extensions going beyond the SM (BSM), the Minimal Supersymmetric extension of the SM (MSSM) predicts two charged Higgs bosons, H^±, and three neutrals: h corresponding to the observed one discovered in 2012 and H and A being additional with respect to the SM. The MSSM phenomenology motivates the focus on events containing a di-τ pair. The MSSM H/A → ττ analysis is thus the core of this thesis. In this thesis, the search for H and A is performed on the data collected with the CMS detector from 2016 to 2018 on proton collisions at a center-of-mass energy of 13 TeV, corresponding to an integrated luminosity of 137 fb^-1. Jets are complex physics objects obtained in the proton collisions occurring at the CERN LHC. Their calibration by the study of events containing a photon and a jet (γ+jet) and the corresponding results on 2018 data are introduced. These results are used in the official CMS jet calibration. No significant deviation above the expected background is observed in the MSSM H/A → ττ analysis. Model-independent limits are then set on the product of the cross section and branching fraction for the production via gluon-fusion or in association with b quarks. These limits range from 15 pb at 110 GeV to 3x10^-4 pb at 3,2 TeV for gluon-fusion and from 1,2 pb at 110 GeV to 3x10^-4 pb at 3,2 TeV for b-associated production. In the M¹²⁵_h scenario, these limits translate into an exclusion region in the (m_A, tan β) plane. Values of m_A below 600 GeV are excluded and this goes up to 2 TeV for tan β >≅ 50. In the M¹²⁵_H1 (CPV) scenario, the region is given in the (mH^±, tan β) plane. Values of m_H^± below 400 GeV are excluded and this goes up to 1,4 TeV for tan β ≅ 20. To test any theory involving Higgs or Z bosons decaying to τ⁺τ^-, the reconstruction of di-τ mass in a faster and more accurate way than the existing methods is crucial. However, it is an arduous task due to existence of neutrinos as decay product of each τ lepton which are invisible to detectors at LHC. Machine learning techniques bring a solution for this task. The reconstruction of the di-τ mass by a deep neural network (DNN) is achieved in this thesis from 50 GeV to 800 GeV with a 20% resolution at 50 GeV, 26% at 250 GeV and 22% at 800 GeV. This DNN is 60 times faster and better at describing the Z boson than the SVFIT algorithm currently used in CMS. (author)