Regression Deep Neural Networks for top-quark-pair resonance finding in the dilepton channel: feasibility study and foreseen improvements over traditional analysis techniques with the ATLAS experiment at the LHC

The ATLAS experiment at the LHC has the potential to find evidences of new physics beyond the Standard Model (BSM). Some BSM theories predict the existence of new massive particles decaying to pairs of top quarks ($t\bar{t}$), that can be detected by ATLAS. The key distribution to consider to prove the existence of such hypothetical particles is the invariant mass of the reconstructed $t\bar{t}$ pair, $m_{t\bar{t}}$. The reconstruction of such an invariant mass from the detected particles in each event is particularly hard in the dilepton $t\bar{t}$ decay channel, where the final states include two neutrinos, which always escapes undetected. Several traditional analysis techniques already exist, but a new data analysis tool could improve the accuracy of the $m_{t\bar{t}}$ estimate. This thesis consists of a feasibility study, in which deep learning, an innovative multivariate analysis technique, is applied to simulated ATLAS events to perform a regression task. A deep neural network is successfully implemented in the feed-forward architecture and it proves to be a useful tool for reconstructing the invariant mass of top-quark pairs decaying in the dilepton channel. Comparing this innovative method with some of the most commonly used traditional analysis techniques, it was possible to show significant performance improvements in both the SM and BSM scenarios.