Ceschia, Matteo; (2024) Application of Deep Learning Techniques for Identification of Double Beta Decay Events in the SuperNEMO Experiment. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Matteo_Ceschia_PhD_Thesis_MatteoCeschia_FINAL.pdf - Other Download (17MB) | Preview

Abstract

Neutrinoless double-beta (0νββ) decay is a hypothesised lepton number violating process which has not been observed yet. Its lifetime of more than 1026 years makes it one of the longest decay processes in the universe. If detected, it would shed a light on beyond the Standard Model physics, and in particular on the nature of the neutrino and on the origin of matter-antimatter asymmetry. The SuperNEMO demonstrator detector aims to reach a half-life sensitivity of 5.85 × 1024 years with a total exposure of 17.5 kg × years with 82Se as source (corresponding to ⟨mββ⟩ < 0.2 − 0.55 eV) [1]. Its unique tracker-calorimeter design allows full topology and kinematics reconstruction, providing the opportunity to study both single- and double-electron energy distributions. A scaled up version of SuperNEMO will be able to probe the Majorana neutrino masses down to the ∼50 meV level with the potential to explore the underlying lepton number violating physics mechanism that is not accessible with other experimental techniques. First, this thesis studies the different behaviour of the modelling of particle tracks in the presence of the magnetic field induced by SuperNEMO magnetic coil. This analysis shows the presence of edge effects on the main detector walls and highlights a reconstruction artefact that gives rise to an excess of events on veto walls. Second, the application of deep-learning based methods is investigated for the first time in the SuperNEMO experiment. A convolutional neural network (CNN) with an encoderdecoder architecture was designed in order to recover the non-responsive tracker cells from images of partial tracks. Its performance is compared to classical algorithms with respect to their ability to retrieve missing cells. The 0νββ identification rate of the proposed deep-learning algorithm is improved by ∼ 6% with respect to traditional algorithms. Moreover, the CNN is shown to recover malfunctioning cells with a significantly lower (up to a factor of 5) false positive rate compared to a benchmark cut-flow based analysis. Lastly, the thesis describes the application of deep learning in order to improve the classification of events. A CNN is designed to improve the classification of ββ events using all the available detector information. It is shown that the CNN achieves a better performance in the selection of 0νββ, with a higher background suppression rate with respect to the traditional SuperNEMO reconstruction algorithm. In particular, for the case of the 2νββ decay being the sole background to 0νββ, this method achieves a sensitivity of T 0ν 1/2 > 8.0 × 1024y at 90% CL, which is 36% higher than that achieved by standard reconstruction and analysis methods. This is the first use of deep-learning techniques in the SuperNEMO experiment, and it is expected to pave the way for future applications, widening its scope to background events and energy reconstruction.

Download activity - last month

Export as CSVXMLJSON

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as JSONCSVXML

Archive Staff Only

View Item