Graph Building for Graph Neural Networks for Photon Reconstruction in the Belle II Calorimeter

This thesis presents the implementation and evaluation of the GCN algorithm for photon energy reconstruction in the Belle II electromagnetic calorimeter. The GCN algorithm, a machine learning technique based on graph convolutional networks, uses fuzzy clustering. The model is employed to analyze and reconstruct photon energies of single photon events using Monte Carlo generated and simulated data. This study focuses on the graph-building process and on investigating the graph input features utilized in the GCN model to optimize the algorithm’s performance. In particular, the node input features and the number of edges in a graph as well as the edge weights are studied. To assess the effectiveness of the GCN algorithm, a comparison is made with the algorithm of the Belle II Analysis Software Framework and the GravNet algorithm. The results show that the GCN model outperforms the Belle II Analysis Software Framework model by up to 15 % in the photon energy range of 0.01 GeV to 2.5 GeV while maintaining comparable performance to the GravNet model. For photon energies exceeding 2.5 GeV, the GCN model performs better than the GravNet model, with an improvement of up to 16 %.