Determination of missing transverse momentum using multivariate methods for a differential top pair production cross section measurement at CMS

In this thesis, a deep neural network (DNN) is developed, aiming to correct for detector effects in the reconstruction of observables, which are used in the binning of a differential top pair ($t$$\bar{t}$ production cross section measurement. These binning variables are given by the momentum imbalance in the plane transversal to the beam axis ($\mathrm{p}_{T}^{miss}$) and the azimuthal angle between $\mathrm{p}_{T}^{miss}$ and the lepton closest to it (|$\bigtriangleup$$\psi$($\mathrm{p}_{T}^{miss}$), nearest $\iota$|). The DNN is trained based on MC simulations corresponding to the data recorded at the CMS experiment in Run 2 of the LHC, with a center-of-mass energy of 13 TeV and total integrated luminosity of 138 fb-1.
The input features of the DNN are validated by performing goodness-of-fit tests of the feature distributions between data and simulation. Subsequently, the settings of the DNN are optimized in a Bayesian search. The performance of the optimized DNN is then studied. The resolution and bias of $\mathrm{p}_{T}^{miss}$ and |$\bigtriangleup$$\psi$,($\mathrm{p}_{T}^{miss}$, nearest $\iota$)| are compared for reconstruction with and without application of the DNN-based correction in simulated samples of the $t$$\bar{t}$ process and its backgrounds. Additionally, the event migrations corresponding to the reconstruction of the $t$$\bar{t}$ binning variables are studied for reconstruction with and without the DNN.
The studies showed that by applying the DNN-based correction, the reconstruction resolution of the binning variables could be increased significantly, resulting in less event migrations in the binning of the differential $t$\bar{t}$ cross section measurement.