Tau Lepton Classification With Graph Neural Networks

Abstract

This thesis investigates the use of deep learning to classify tau leptons in high- energy physics experiments by employing Graph Neural Network (GNN)s. Tau leptons, unstable subatomic particles of the lepton class, decay rapidly and produce tau jets. These jets must be distinguishable from the more common quark jets in particle collisions. Given the complexity of separating these particles in the ATLAS experiment at the Large Hadron Collider (LHC), this thesis proposes using GNNs as a novel approach for more accurate identification within this context. The GNNs developed in this thesis are trained and evaluated on the same data as the current Recurrent Neural Network (RNN) model for tau jet classification. In the final analysis, the GNNs are evaluated against the Recurrent Neural Network (RNN) model to assess their effectiveness. The thesis explores three main GNN architectures: Graph Convolutional Networks (GCN), Graph Attention Networks (GAT), and Dynamic Graph Convolutional Neural Networks (DGCNN), and also proposes a more complex Multi-modal Graph Neural Network (MmGNN) design. The findings in this thesis demonstrate that DGCNN and GAT outperform the RNN model on 1-prong data. However, in the more complex data in the 3-prong dataset, the DGCNN is marginally better than the RNN. These results under- score the potential of GNNs to enhance the accuracy of tau jet identification.