Improvement of the top-quark reconstruction for the top-quark mass measurement using machine learning techniques

The measurement of the top-quark mass is important for precision tests of the Standard Model. The precision of the $m_{top}$ measurement is limited by systematic uncertainties. A measurement of $m_{top}$ in the lepton+jets channel at $\sqrt{s}$ = 8 TeV was performed by the ATLAS experiment using a boosted decision tree (BDT) for the event selection optimisation. The BDT was used to reject wrongly reconstructed events in order to reduce the systematic uncertainty of the top-quark mass. In the thesis presented here, investigations on the event selection and reconstruction for the $m_{top}$ measurement at $\sqrt{s}$ = 13 TeV are performed using simulated events. These investigations are achieved using both a BDT similar to the 8 TeV analysis as well as using a deep neural network (DNN), which is a more sophisticated machine-learning technique. In comparison to the BDT selection at 8 TeV, the matching efficiency of the BDT selection at 13 TeV is slightly lower but leads to a larger purity. It is also found that variables with a small separation can be removed from the training without decreasing the BDT performance significantly. First studies with the DNN improve the purity of the event selection compared to the BDT selection and lead to a better $m_{top}$ resolution. The DNN training therefore provides a new classifier of promising potential for the full measurement of $m_{top}$.