Event Reconstruction in SNO+ with Graph Neural Networks
Loading...
Authors
Cheng, Sabrina Zhongliang
Date
2024-06-05
Type
thesis
Language
eng
Keyword
SNO+ , Neutrino Physics , Machine Learning
Alternative Title
Abstract
SNO+ is a large-scale neutrino experiment with the primary goal of searching for neutrinoless double beta decay in Te-130. The reconstruction of the position of particle interaction events in the detector volume is central to the SNO+ data analysis pipeline.
This thesis presents a new method of position reconstruction using the machine learning method of Graph Neural Networks (GNN).
Based on a simulated dataset of events in the energy range of 0.5 to 5 MeV, a comparison is made between the GNN, the current reconstruction method, and an independent convolutional neural network method developed within the SNO+ collaboration. The GNN reduces the average fit error by 13.1% over the current reconstruction method, and by 12.8% over the convolutional neural network method.
In addition, the GNN reduces the percentage of events misplaced outside of the detector from 6% to 2% compared to the current method of reconstruction. A separate GNN is trained to reconstruct energy using the same architecture as the position reconstruction model to test the ability of the GNN to generalize to other applications. Under these conditions, GNN improves the average fit error by 16.8% over the current method of energy reconstruction. When operating on a CPU, the GNN has a comparable inference speed to the current reconstruction method, and when operating on a GPU, the GNN's inference speed is 96.7% faster.