Adversarial neural networks for associated top pair and Higgs Boson production in the di-photon decay channel
Takeva, Emily Petrova
The main topic of this thesis is classification with Adversarial Neural Networks, which are for the first time used in an analysis targeting final states in which the Higgs boson decays to pairs of photons (H → γγ). The analysis uses 139 fb⁻¹ of proton-proton collision data recorded at √s = 13 TeV by the ATLAS experiment at the Large Hadron Collider, and targets the associated top pair and Higgs boson production (ttH). Backgrounds with non-resonant photon pairs such as multi-jet or top-antitop pair production in association with photons are mainly rejected by using the photon kinematics. The signal is extracted from a fit of the di-photon invariant mass (Mγγ) distribution, which consists of a narrow signal peak on the top of a substantial background. Using the kinematic variables of the photons for the classification causes the background Mγγ distribution to peak at the Higgs boson mass value, due to these variables being correlated with Mγγ. This sculpted background distribution is hard to parametrise with a simple functional form needed for the background fit to the Mγγ sidebands. The novel adversarial neural network approach developed in this thesis enables designing a classification discriminant independent of Mγγ, which removes the sculpting, while keeping the classification efficiency optimally high. Additionally, work towards the creation of a programme to deal with the interpolation of energy values used for the incident single particles in the fast calorimeter simulation of the ATLAS experiment, which is used to date, is also presented in this thesis.