Classification of particles with a convolutional neural network for neutrino and antineutrino events in the NOVA experiment

Abstract

We are in the precision era of neutrino oscillation experiments. The NOvA experiment newlineaims to precisely determine neutrino oscillation parameters (and#948;CP , and#952;23, and newlinesign of and#916;m2 newline32). High energy physics experiments like NOvA produce vast amounts newlineof data. This makes traditional analysis methods challenging and time-consuming. newlineMachine learning algorithms excel at extracting patterns and insights from large newlinedatasets, aiding in the identification of relevant particle signatures, event classification, newlineand background noise reduction. This makes using machine learning in newlinehigh energy physics very beneficial. newlineNOvA uses machine learning to develop various tools to assist the analysis. For newlineexample, Event CNN is used to to classify candidate neutrino interactions. Identification newlineof final state particles of an event is done through Prong CNN. Prong newlineCNN is trained to identify all the final-state particles (e±,p+, and#956;, and#960;±, and#947;) of a given newlineneutrino event. newlineIn this thesis, we gave detailed description of NOvA experiment and various algorithms newlineused by NOvA for simulation and reconstruction. Then, we described newlinethe development of Prong CNN, and modifications made compared to previous newlineversion. We trained single Prong CNN using both neutrino and anti-neutrino newlineevents. In our updated Prong CNN, we were able to reduce runtime on CPUs newlinewithout compromising performance. Our Prong CNN is now more computationally newlineefficient, enabling higher classification efficiency within fewer training epochs. newlineThe newer prong CNN has a global efficiency greater than 86% with improved the newlineclassification efficiency by 3% compared to previous Prong CNN. newlineBesides developing the prong CNN for the NOvA experiment, a phenomenological newlinestudy was conducted within the framework of the 3+1 sterile neutrino model using newlineGLoBES. The study was aimed to assess the impact of sterile neutrinos on the newlinedegeneracy resolution capabilities of both the NOvA and DUNE experiments. newline