Machine learning in NO$\nu$A near detector vertexing

Abstract

This work presents an alternative method for a reconstruction machine learning algorithm in place of the one currently used by the NO$\nu$A (NuMI Off-axis $\nu_e$ Appearance) group for the analysis of neutrino events at the Fermilab near and far detectors. Current reconstruction methods are flawed because there are many challenges associated with finding the proper vertex of a neutrino event including background noise, incorrect prong creation, and secondary vertecies. This work presents a regression-based convolutional neural network (CNN) that analyzes 2-dimensional pixel maps from NO$\nu$A ’s catelog of forward horn current (FHC) and reverse horn current (RHC) h5 files and more accurately predicts the location of the vertex for each coordinate direction. Additionally, this model can be implemented into NO$\nu$A using their primary framework, NO$\nu$ASOFT and applied to larger data sets to get a better comparison, or even to expand the model to take on secondary vertexing.