Multi-plane Neural Networks for Event Reconstruction in LArTPCs



Henzerling, Jaggar
(2022) Multi-plane Neural Networks for Event Reconstruction in LArTPCs. Unspecified thesis, University of Liverpool.

[img] Text
201277823_Mar2022.pdf - Unspecified

Download (46MB) | Preview

Abstract

With the advent of high precision Liquid Argon Time Projection Chambers (LArTPCs), the need for fast and efficient data analysis has grown exponentially. The Shorte Baseline Neutrino (SBN) program [17], consisting of MicroBooNE, ICARUS, and SBND and located along the Booster Neutrino Beam (BNB) at Fermilab aims to collect millions of neutrino events across its experimental lifetime. The high statistics and and rate of data acquisition serves to elucidate certain mysteries about the neutrino, such as the existence of a sterile flavour, but comes with several challenges. Primary concerns such as the high flux of cosmic rays and the difficulty in classifying certain particle types in the detector add to the need for fast and automated data analysis software to discriminate between neutrino event varieties. Machine Learning (ML) and specifically Convolutional Neural Networks (CNN’s) offer a unique and attractive option towards event reconstruction in LArTPCs. Neural networks, modelled after the connections between neurons in brains, learn deep and often complex features based on data supplied during training. With the firm backing of simulation software such as GENIE [36], training data driven by simulation and experiment is abundantly available. While perfect translation of information from experiment to simulated data is difficult, such difficulties in the data generation domain is left for future work. By utilising these generated datasets, one can train neural networks to predict energies, particle types, interaction modes, and even segment separate instances of particles in the pursuit of reconstructing a neutrino event. By reconstructing the contents in a LArTPC through this computational tool to determine event topologies, evaluation of different physical models can be tested quickly and automatically at SBN. This thesis aims to apply Sparse Submanifold Convolutions to create network architectures suited for LArTPC data with minimal memory footprints in GPUs. Furthermore, contrary to some contemporary studies, this thesis focuses on the application of feature fusion in order to co-learn features across multiple 2D planes. This approach, contrary to using 3D (full detector volume) data, takes inspiration from the structural output of wire-based LArTPC detectors, and serves to demonstrate the efficacy of deep learning techniques over single-plane 2D methods on individual wire planes. These feature fusion models, deemed multi-plane models in this work, demonstrate increased classification accuracy on particle types on the order of 90%. Furthermore, cosmic neutrinos have been demonstrated to be removed with 99% efficiency, allowing event selection for further analysis. Simulated events were segmented by semantic classes based on geometry with an overall efficiency of 98%, leading over other 2D models. Segmenting by particle ini stance was determined at an overall clustering efficiency of 97%, or an overall Adjusted Rand Index (ARI) of 0.9. The clustering works well on the majority of particles, but key error modes prevent the current model from applicability in an end-to-end reconstruction chain. Lastly, the tools developed for use in LArTPC physics was applied to the problem of thyroid cancer diagnostics as a separate project. Employing a dense multi-scale (Inceptionstyle) network a 98 % diagnostic accuracy was reported per-slide. Furthermore, a novel voting strategy is proposed which reports a per-patient diagnostic performance of 98%. Neural networks are demonstrated to be valuable tools both in physcs and outside for various classification and clustering tasks, with the goal of creating a larger analysis pipeline for end-to-end reconstruction.

Item Type: Thesis (Unspecified)
Divisions: Faculty of Science and Engineering > School of Physical Sciences
Depositing User: Symplectic Admin
Date Deposited: 29 Jun 2022 13:14
Last Modified: 18 Jan 2023 20:57
DOI: 10.17638/03157242
Supervisors:
URI: https://livrepository.liverpool.ac.uk/id/eprint/3157242