Energy Reconstruction in KM3NeT/ARCA using Graph Neural Networks

Abstract

Neutrinos, fundamental particles that interact only weakly with matter, are crucial to our understanding of both particle physics and a variety of astrophysical phenomena. Despite their importance, detecting neutrinos is a significant challenge due to their elusive nature. Large-scale water Cherenkov detectors, such as those of the KM3NeT experiment which are located deep in the Mediterranean Sea, are designed to capture the Cherenkov light produced when high-energy neutrinos interact with atomic nuclei in water. These interactions allow physicists to probe the properties of neutrinos and study extremely energetic astrophysical objects and processes. One of the key challenges in these experiments is accurately reconstructing the energy, direction and other characteristics of each neutrino event from the data collected by the detector optical sensors. Traditionally, event reconstruction has relied on classical algorithms, such as maximum likelihood estimation, however the rise of modern Deep Learning methods as powerful means of extracting information from data has demonstrated the potential to improve upon these techniques. In particular, Graph Neural Networks (GNNs) have emerged as a promising approach due to their ability to naturally incorporate the non-grid-like structure and sparsity of neutrino event data. This thesis investigates the application of DynEdge, a particular GNN architecture, to energy reconstruction in the KM3NeT/ARCA detector, using the GraphNeT Deep Learning framework, developed by the IceCube neutrino experiment. In this context, the KM3NeT Deep Learning data format was integrated into GraphNeT, allowing for the first use of this tool in KM3NeT and facilitating its future use within the collaboration. Various configurations of models and training setups are evaluated on two versions of our neutrino event dataset, snapshot and triggered data, in order to determine the optimal approach to energy reconstruction. The results allow for a direct comparison between DynEdge and ParticleNet --the GNN model previously used by KM3NeT-- validating the performance of both models. This work also presents significant improvement in the energy reconstruction of low-energy neutrino events, achieved by the implementation of event weighting to balance the energy distribution of the input dataset. This strategy also contributes to the elimination of biases in the training process, leading to a more generalized model. Finally, the thesis explores the challenges of reconstructing high-energy (PeV scale) neutrino events, providing insights into the predictive limitations of the model, which stem from the physical constraints that apply at this energy scale.