Studien zur Rekonstruktion der invarianten Masse von Top-Antitop-Paaren im dileptonischen Kanal mit Hilfe eines neuronalen Netzwerkes am ATLAS Detektor.

The Standard Model (SM) of particle physics forms the foundation of our understanding of the most fundamental forces and particles. Nevertheless, phenomena such as dark matter or the observation of neutrino-oscillation provide good reasons to believe that the SM is not a complete theory. Precision measurements of the SM in collision experiments are crucial for identifying deviations that could point to new physics. The studies in this thesis are motivated by the measurement of the top quark Yukawa coupling $Y_t$ relative to the SM value in the dileptonic decay channel of top-antitop ($t\bar{t}$) pairs at the ATLAS experiment. The used (simulated) data correspond to the proton-proton collision data recorded between 2015 and 2018 at $\sqrt{s}=13$ TeV with an integrated luminosity of 140 fb$^{-1}$. In the dileptonic channel, the $t\bar{t}$ pair decays into two $b$ quarks and two $W$ bosons, which then decay into one lepton and neutrino each. Virtual corrections through the exchange of a Higgs boson between the produced top quarks modify the spectrum of the invariant mass of the top-antitop pair $m_{t\bar{t}}$ in the region of the production threshold ($\approx2m_t$). A precise analysis of this distribution provides insights into the strength of the top-Yukawa coupling. Due to the two neutrinos in the final state, the reconstruction of $m_{t\bar{t}}$ is only possible through approximation methods that incorporate kinematic constraints on the $W$ mass and top quark mass as well as the reconstructed missing transverse momentum. Therefore, a neural network using techniques from the field of deep learning is motivated and presented in this thesis, which reconstructs $m_{t\bar{t}}$ through regression using high-level observables as input. As inputs, the invariant masses $m_{e\mu}$, $m_{eb_1}$, $m_{eb_2}$, $m_{\mu b_1}$, $m_{\mu b_2}$, and $m_{b_1b_2}$ of the electrons, muons, and $b$-jets in the final state, together with the magnitude of the missing transverse momentum $E_\text{T}^\text{miss}$, are used. The output distribution of the neural network is analyzed for datasets generated by pythia8 and Herwig7 to investigate the influence of different $t\bar{t}$ modeling. Additionally, the agreement between simulation and experimental data is examined. A maximum-likelihood fit for a future measurement of $Y_t$ is presented. Here, the experimental data and the expected $t\bar{t}$ signal for different values of $Y_t$ are compared, and the best agreement is given as an estimate for $Y_t$, considering systematic uncertainties. Since the systematic uncertainties were not implemented due to time constraints, an initial fit based on Asimov data is performed to estimate the measurement precision. Compared to the easily reconstructed invariant mass $m_{bbll}^\text{reco}$ of the leptons and $b$-jets, the distribution from the neural network shows an approximately 3.5% lower uncertainty in the estimation of $Y_t$ (excluding systematics). The systematic uncertainty of the $t\bar{t}$ modeling differences by Pythia8 and Herwig7 is also examined.