Classification of Top-Antitop Events in Association with Bosons Using Graph Neural Networks in Atlas

At the ATLAS experiment at CERN, proton-proton ($pp$) collisions of an energy of $13.6$ TeV produce complex final states containing a wide range of overlapping physical processes and share similar final state particles. This complexity creates major challenges to clearly distinguish rare signal events accurately. Additionally, limitations pertaining to the detector and uncertainties in particles' reconstruction lead us to the applications of sophisticated event classification techniques. This thesis presents a method applying a graph neural network (GNN) to the classification of final states containing a $t\bar{t}$ paired with additional particles collectively referred to as $t\bar{t} + X$, wherein $X$ can represent a Higgs boson ($H$), a $Z$ or $W$ boson, or a photon ($\gamma$), using Monte Carlo simulated events. In contrast to standard techniques like graph convolutional neural networks, including those requiring information to be in grid-like structures, GNNs give options to inherently encode the physically relational and spatial nature of a particle on a per-event basis, representing particles and their interactions as nodes and edges of graphs. Utilizing an application of message passing amongst nodes of graphs, GNNs give a richer description of event topologies, thus aiding in a richer event identification. A series of model architectures are studied, which includes Graph Attention Network and Graph Transformer, to find the best version to achieve an optimization of multi-class classification across a spectrum of event types.