engRieger, MarcelErdmann, MartinSchmidt, AlexanderSearch for Higgs boson production in association with top quarks and decaying into bottom quarks using deep learning techniques with the CMS experimentinfo:eu-repo/classification/ddc/530CERNCMSHiggs bosonTop quarkdeep learningmachine learningneural networksparticle physicsAfter the discovery of the Higgs boson during the first run of the Large Hadron Collider (LHC), precise measurements of its properties and couplings to other particles are conducted by the ATLAS and CMS experiments. Due to its high mass, the Standard Model of particle physics (SM) predicts that the top quark, among all fermions, exhibits the strongest coupling to the Higgs boson with a Yukawa coupling constant close to unity. Therefore, the accurate measurement of Higgs-top coupling is essential for probing predictions of the SM and to constrain theories that reach beyond it. This thesis presents a search for the associated production of a Higgs boson with a top quark pair at a center-of-mass energy of $\sqrt{s}$ = 13 TeV. The data from proton-proton collisions of the second run of the LHC was recorded by the CMS detector in 2016 and corresponds to an integrated luminosity of 35.9 fb$^{-1}$. The analysis focuses on events where the Higgs boson decays into a pair of bottom quarks and the decay of the $t\bar{t}$ system involves one or two electrons or muons with opposite charge. In addition to the lepton requirements, events are selected to have missing transverse energy and at least four energetic jets of, which at least three must have a $b$ tag. The analysis is dominated by systematic uncertainties in the normalization of $t\bar{t}$ background contributions with additional heavy-flavor jets, especially from $t\bar{t}\!+\!b\bar{b}$, $t\bar{t}\!+\!b/\bar{b}$, $t\bar{t}\!+\!2b$, and $t\bar{t}\!+\!c\bar{c}$ production. To reduce their impact on the analysis, a novel event categorization scheme based on deep neural networks is introduced. In contrast to binary classification, the networks perform a multi-class classification, which attributes events with a probability to originate from either the signal process or a particular background process. The information of the highest probability is exploited to create mutually exclusive categories that are enriched with events of one of the considered physics processes. As the $t\bar{t}+$heavy-flavor normalization uncertainties are treated as uncorrelated in the signal extraction procedure, the simultaneous isolation of backgrounds using this approach leads to more accurate background constraints and, in turn, to a gain in signal sensitivity. Results are obtained in a simultaneous measurement over 24 orthogonal categories defined by the lepton channel, the number of jets, and the most probable process as attributed by the deep neural networks. In each category, the distribution of the network output unit that corresponds to the assigned process category is fitted to data. The best fit value of the $t\bar{t}H$ signal strength modifier $\mu\,=\,\sigma / \sigma_\text{SM}$ is extracted in a profile likelihood fit with systematic uncertainties incorporated as nuisance parameters. Upper limits on $\mu$ are computed using the asymptotic CL$_\text{S}$ method at 95% confidence level. The SM predicts a $t\bar{t}H$ production cross section of $\sigma_{SM}\,=\,507\,^{+35}_{-50}$ fb with NLO QCD accuracy and NLO electroweak corrections. This analysis measures an observed upper limit on the $t\bar{t}H$ signal strength modifier of $\mu_\text{obs} < 1.85$, which excludes greater values with 95% confidence given the measured data. The background-only hypothesis is disfavored due to an expected limit of $\mu_\text{exp}\,<\,0.99\,^{+0.42}_{-0.29}$ in case of a signal strength as predicted by the SM. The observed best fit value is measured as $\mu_\text{obs}\,=\,0.98\,^{+0.50}_{-0.48}\,=\,0.98\,^{+0.26}_{-0.25}\,(\text{stat.})\,^{+0.43}_{-0.41}\,(\text{syst.})$. A value of $\mu_\text{exp}\,=\,1.00^{+0.55}_{-0.49}$ is expected from simulation. The corresponding observed (expected) significance amounts to 2.04 (1.92) standard deviations above the expected background. The total uncertainty predominantly originates from systematic effects. The analysis in the semi-lepton channel as presented in this thesis contributed to the first observations of both $t\bar{t}H$ production and Higgs bosons decaying into bottom quarks, performed with the CMS experiment.Aachen 1 Online-Ressource (178 Seiten) : Illustrationen, Diagramme (2019). doi:10.18154/RWTH-2019-06415 = Dissertation, RWTH Aachen University, 2019info:eu-repo/semantics/doctoralThesisinfo:eu-repo/semantics/publishedVersion2019info:eu-repo/semantics/openAccessDEhttps://publications.rwth-aachen.de/record/763526https://publications.rwth-aachen.de/search?p=id:%22RWTH-2019-06415%22StudentsStudent Financial Aid ProvidersTeachersResearchersinfo:eu-repo/semantics/altIdentifier/doi/10.18154/RWTH-2019-06415