Application of Convolutional Neural Networks in Neutrino Physics

Abstract

Abstract Convolutional neural networks (CNNs), as a deep learning algorithm, have successfully been used for analyzing visual image data over the past years. As some of the physical experiments can produce image-like data, it is more than fitting to combine the interdisciplinary knowledge between high energy physics and deep learning. Especially in the domain of neutrino physics, the particle classification problem has played an important role and CNNs have shown exceptional results for the image classification. In this paper, results of application of CNN called SE-ResNET on Monte Carlo simulated image data is presented. These visual images are tailored to fit measured data from future Deep Underground Neutrino Experiment. The image classification focuses primarily on neutrino flavor classification, namely on classification of charged current (CC) electron νe , CC muon νμ , CC tauon ντ and neutral current (NC); and secondarily on other characteristics of the image, such as whether the observed particle is neutrino or antineutrino. The results are important for further physical analysis of the neutrino experiment event, e.g. for study of neutrino oscillation.