Neutrino interaction vertex reconstruction and particle identification in the MicroBooNE detector
This thesis presents the results of a study measuring and improving the quality of neutrino interaction vertex reconstruction and particle identification (PID) in the MicroBooNE detector. The detector comprises a liquid argon time-projection chamber (LArTPC) with a light-collection system, permitting precise tracking of neutrino interaction final states. MicroBooNE's primary physics goal is to resolve the low-energy electron neutrino appearance anomalies observed at MiniBooNE and LSND. The experiment therefore requires high-quality neutrino interaction vertex reconstruction and PID, which together strongly influence event reconstruction quality and energy/momentum estimation. Improvements to the vertex reconstruction are made through the development of powerful new variables and the application of machine learning techniques; these algorithms are now the default used at MicroBooNE and have enabled new studies of neutrino interactions with up to six charged particles in the final state. A robust PID method (FOMA) is developed using a novel analytic approximation to the mode of the dE/dx distribution. A deep learning PID method (PidNet) is also proposed, based on convolutional neural networks (CNNs) and a semi-supervised representation learning method. The performance of the two approaches is compared and contrasted with PIDA, the default PID algorithm used at MicroBooNE. This work concludes by assessing the impact of the tools and methods developed in this work on particle energy estimation in MicroBooNE.