Deep Learning techniques for the observation of the Higgs boson decay to bottom quarks with the CMS experiment

The Higgs boson was discovered at the CERN LHC by both the ATLAS and CMS collaborations in 2012 with a mass near 125 GeV. The characterization of the newly discovered particle has been one of the principal goals of the LHC experiments since. The main result reported here marks an important step in the effort of characterizing the Higgs boson: this thesis describes the first observation of the H→bb̄ decay with CMS data. The measurement of the H→bb̄ decay mode probes directly the Yukawa coupling of the Higgs boson to down-type quarks. Consequently, it is a fundamental test of the mechanism that generates the masses of the fermions, and of the consistency of the Higgs properties with the Standard Model hypothesis. The H→bb̄ decay observation comes after the bosonic decay modes of the Higgs, $\gamma\gamma$, ZZ, and WW, and the fermionic decay into $\tau$$\tau$ were all firmly established. The tt̄H production mode has also been observed, thus probing directly the coupling to up-type quark. The H→bb̄ observation is therefore closing a chapter in the Higgs Physics at the LHC: with all the most accessible production and decay modes now observed, the focus is shifting to rare decay modes, precision measurements and differential cross-section measurements. The analysis presented uses 2017 data collected by the CMS experiment at $\sqrt{s}$ = 13 TeV, which corresponds to an integrated luminosity of 41.3 fb$^{–1}$ . The vector boson associated production mode (VH) with 0, 1 or 2 charged leptons in the final state is targeted, as it’s the most sensitive for the H→bb̄ decay. The Higgs boson signal is extracted via a likelihood fit and an excess of 3.3 standard deviations over the background-only hypothesis is measured (with 3.1 standard deviations expected), corresponding to a signal strength μ = 1.08 ± 0.34. The analysis is combined with previous results for VH(bb̄), reaching 4.8 (4.9 expected) standard deviations for the VH(bb̄) process, with a signal strength of 1.01 ± 0.22. In combination with results targeting different production modes, namely the VBF H(bb̄) analysis using Run 1 data, the tt̄H(bb̄) with 2016 data and the inclusive search for H(bb̄) in the boosted regime, a significance of 5.6 (5.5 expected) standard deviations is reached, corresponding to a signal strength of 1.04 ± 0.20. The heavy usage of Deep Learning techniques that I largely developed in my Ph.D. work was a crucial element for the observation of the H→bb̄ decay. Four different deep neural networks have been used: for tagging b jets, which are the reconstructed objects originating from the H→bb̄ decay; for the calibration of their energy and momentum; for background classification in control regions; and for discriminating the signal from the backgrounds. Machine Learning already played an important role in the previous searches for VH(bb̄), but with 2017 data Deep Learning was introduced and very quickly became fundamental. Deep Learning techniques are now becoming more and more important at the LHC, not only at the analysis level, but also because they are starting to be an integral part of the reconstruction algorithms in CMS. Hence, an important part of the thesis is dedicated to Deep Learning techniques, and their application to the b jets is shown as a use case. The thesis is structured as follows: the Standard Model framework, with a focus on the Higgs mechanism, is described in Chapter 1. Also, a summary of the most important results achieved at the LHC on the properties of the Higgs boson is given. Chapter 2 is dedicated to the experimental apparatus: after a description of the LHC machine, the most important features of the CMS detector are presented. The reconstruction of physics objects is performed in multiple steps: lower level objects’ reconstruction is included in Chapter 2. Higher level objects which are then used in the analysis are described in Chapter 3. Chapter 4 introduces Deep Learning concepts whose application is present both in Chapter 5 and Chapter 6. Chapter 5 is dedicated exclusively to Deep Learning applications: the object under study are the b jets, which are produced at the LHC both in the H→bb̄ decay and in a number of background processes. Two tasks are important in analyses with b jets in the final state: the correct reconstruction of the jet momentum and the ability to separate b jets from jets originating from gluons, light quarks and charm quarks. We usually talk about "b jet energy regression" for the calibration of the jet transverse momentum and "b tagging" for the discrimination of b jets from the other hadronic jets. The b jet energy regression used in the VH(bb̄) analysis is described in detail. Subsequently, a Deep Learning based b tagging algorithm, which, unlike most of the tagging algorithms uses only reconstructed tracks but no reconstructed secondary vertices, is presented. The algorithm, called "DeepVertex", exploits the ability of Deep Neural Networks to learn from raw data, aiming to infer the secondary vertex properties from tracks and clusters of tracks in the hidden layers of the network. My work focused on the development of the regression Deep Neural Network in parallel with the ETH group searching for di-Higgs production, then on the validation with data of the trained model for the VH(bb̄) analysis and potentially for the CMS collaboration. I also carried out the development and optimization of DeepVertex in simulation, which has now reached results useful for the entire CMS collaboration and is ready for deployment in data. Chapter 6 covers the VH(bb̄) analysis with 2017 data and the combination with previous analyses. My first contribution to the analysis was the aforementioned b jet regression and its validation. Subsequently, I worked on the optimization and the inference of the Deep Neural Networks used in the multivariate analysis together with other members of the analysis team. The analysis relies heavily on Deep Learning, both for signal discrimination and to isolate background sources, thus improving the background modeling. The outlook for H→bb̄ and conclusions are in Chapter 7. In this last chapter, a preview of the search for the Higgs boson decay into muons using the full Run 2 data collected by CMS is also presented. The H → μμ decay is important as it’s the most viable channel to probe the decay to the second generation of fermions at the LHC. I was involved in the search for H→ μμ in the VBF production channel, and in particular in the optimization of DNN discriminators, thanks to my previous experience. The analysis uses a similar strategy as VH(bb̄). Deep Learning techniques similar to the ones applied in VH(bb̄) turned out to be the best solution to maximize the sensitivity.