Abstract:
Machine Learning (ML) techniques have a long history of applications in High Energy Physics (HEP) experiments, such as at the Large Hadron Collider (LHC), which generates vast amounts of complex data. In this project, I explore the utility of graph-based ML methods, specifically Graph Neural Networks (GNNs) to develop novel data interpretation methodolo- gies. GNNs leverage relational inductive bias by modelling data as graph structures. This offers a more expressive and adaptable data representation than conventional ML techniques that rely on image or sequence data. The key challenges of implementing a GNN for a par- ticular task lie in constructing meaningful graph representations from LHC data, designing effective GNN architectures, and optimising these models for improved performance. This thesis focuses on two applications of GNNs in HEP: (i) event classification and (ii) track-based analysis. In event classification, GNN-based framework is developed to distinguish signals from background events and the performance is compared against Deep Neural Networks (DNNs). And the impact of different graph structures, GNN operators, and architectural designs is investigated. Furthermore, I have developed a novel GNN algorithm that enhances classification performance by incorporating learnable edge weights. In track-based analysis, I explore the potential of GNNs for processing low-level detector outputs. Particularly their ability to identify the presence of high transverse momentum charged particle from tracker hit information. Results demonstrate the advantages and challenges of GNN-based approaches in event classification and track-related tasks. This research contributes to the growing adoption of geometric deep learning in particle physics and lays the foundation for future applications of GNNs in collider-based experi- ments.
