Photon Reconstruction in the Belle II Calorimeter Using Graph Neural Networks

This thesis presents the implementation and performance of the GravNet algorithm for the photon
energy reconstruction in the Belle II electromagnetic calorimeter. GravNet is a machine learning
algorithm based on the concept of graph neural networks. The Belle II Analysis Software Frame-
work is the currently used reconstruction framework that serves as the baseline for comparison in
several studies. GravNet solves many of the conceptual restrictions that limit the performance
of the traditional reconstruction approach, especially in the presence of high levels of beam
background. The studies in this thesis are considered a first validation and are exclusively based
on Monte Carlo generated and simulated data. The GravNet implementation outperforms the
baseline energy resolutions over a large range of photon energies from 0.01GeV to 3.0 GeV by
up to 20 %. In addition, the studies demonstrate substantial improvements of up to 15 % in the
reconstruction of neutral pions from the invariant mass of two-photon systems. GravNet proves
to be a viable and versatile reconstruction algorithm with a promising outlook for a broad range
of present and future applications.