This thesis covers three different topics: electron identification, a cross-section measurement of top-quark pair production with additional charm jets, and a novel reconstruction method of event topologies using machine learning. The first two topics are performed within the ATLAS experiment using data from Run 2 of the LHC. The last topic is not using ATLAS data but can easily be applied to physics analyses within the ATLAS experiment.

The first part studies electron identification in the ATLAS experiment using different machine learning techniques. In a first study, high-level variables were used with a deep neural network. The background rejection in simulated proton-proton collisions is increased by a factor of up to 5.5 over the current identification algorithm, while keeping the signal efficiencies the same. A better performance can also be observed in a subset of the data recorded in Run 2, corresponding to 2.4 pb^-1. The increased performance in simulated collisions can be further improved by using low-level information like the single cells of the calorimeter and the information of multiple track candidates. Additional improvements of factors between two and three can be achieved.

A measurement of the fiducial cross-section of top-quark pairs with additional charm jets is performed in the second part using the proton-proton collision dataset collected at the ATLAS experiment during Run 2 at sqrt(s)=13 TeV corresponding to 140 fb^-1. The focus is placed on final states with exactly one electron or muon and at least five jets. The final result is obtained by also including final states with two leptons and at least three jets The cross-section is measured to be sigma^fid_tt>=2c = 1.33^+0.28_-0.25 pb for events with at least two additional c-jets and sigma^fid_tt+1c = 6.78^+1.02_-0.82 pb for events with exactly one additional c-jet. These results are compared to the predictions of several generated samples. All samples predict a lower value, though within the uncertainties of the measured cross-sections. Extrapolating the fiducial cross-sections to the full phase space results in cross-sections of sigma^total_tt>=2c = 5.6^+1.2_-1.1 pb and sigma^total_tt+1c = 41.3^+6.9_-6.4 pb for events with at least two c-jets and one c-jet, respectively. These values do not include all-hadronic final states.

Lastly, a new method to reconstruct underlying physics processes is introduced in the Topograph. The reconstruction includes intermediate particles by using prior knowledge of particle physics decays. This is done using either graph neural networks or transformer encoder layers. It not only solves the combinatorics of the final state objects but also directly predicts the kinematics of the intermediate particles. The achieved performance matches that of the state-of-the-art machine learning technique in its application to all-hadronic top-quark pair decays.