Study of photon isolation and search for axion-like particles decaying into collimated photon pairs using machine learning with the ATLAS detector at the LHC

This thesis presents a search for axion-like particles decaying into closely collimated photon pairs, extending the sensitivity, for the first time, to masses below 10 GeV in the context of proton colliders. It is complemented by a study of photon isolation. This work is carried out within the ATLAS experiment at the Large Hadron Collider (LHC) at CERN and is based on data collected between 2015 and 2018 (Run 2), at a centre-of-mass energy of 13 TeV, as well as data from the 2022-2024 period (Run 3), at 13.6 TeV.

The first part of this work contributes to the study of photon isolation performance in the ATLAS experiment. Photons produced directly in hadronic collisions are a valuable probe for studying quantum chromodynamics. It is essential to distinguish them from background sources, which mainly originate from light mesons produced within hadronic jets. To achieve this, calorimetric isolation applies a selection based on the energy deposited around the photon candidate in the detector, as photons originating from jets are typically surrounded by significant hadronic activity.

However, the calorimetric isolation energy of photons is imperfectly modelled in simulations, which are essential for preparing analyses and constructing signal and background models. To address this, corrections are applied directly to the isolation energy distributions in the simulations to bring them into agreement with experimental data. A new methodology is proposed for correcting the Run 2 simulations, offering greater accuracy than previous approaches. These corrections account for the photon energy, the overlap of collisions in the detector (pile-up), the detector region involved, and the photon conversion type, as a significant fraction of photons convert into electron-positron pairs within the detector.

The second part of this work proposes a search for a new resonance in the collimated two-photon decay channel, reaching masses below 10 GeV and thus extending the scope of previous analyses carried out at proton colliders. However, standard photon identification tools, based on the shape of electromagnetic showers in the detector, and isolation processes, based on the surrounding energy deposits, are not suitable for very close photons, whose showers and energy deposits overlap in the detector. To overcome this limitation, machine learning tools have been implemented to develop new identification and isolation criteria specifically adapted to highly collimated photon pairs. This approach extends the potential to search for resonances decaying into photon pairs. Several neural network architectures have been explored: a multilayer perceptron (MLP), a convolutional neural network (CNN), a graph neural network (GNN), and a Set Transformer. The performance of these models is exploited to assess sensitivity to the discovery of a new resonance and to establish expected upper limits on the effective cross-section times the branching ratio.

Finally, in the current context of global warming, research plays a significant role. An inventory of greenhouse gas emissions from French research institutions and CERN is presented. The work carried out at the LPNHE, the host laboratory for this thesis, to quantify its carbon emissions and propose measures to reduce them, is also described.