Improving Application of Bayesian Neural Networks to Discriminate Neutrino Events from Backgrounds in Reactor Neutrino Experiments

TL;DR

This paper applies Bayesian Neural Networks to distinguish neutrino events from backgrounds in reactor experiments, using PMT photoelectron data, and demonstrates improved rejection of certain backgrounds but increased difficulty with fast neutron events.

Contribution

It introduces a BNN approach using PMT photoelectron data for neutrino event classification, improving background rejection over previous methods.

Findings

Better rejection of $^{8}$He/$^{9}$Li and uncorrelated backgrounds.

Increased rejection of certain backgrounds compared to prior work.

Higher fast neutron background acceptance than previous methods.

Abstract

The application of Bayesian Neural Networks(BNN) to discriminate neutrino events from backgrounds in reactor neutrino experiments has been described in Ref.\cite{key-1}. In the paper, BNN are also used to identify neutrino events in reactor neutrino experiments, but the numbers of photoelectrons received by PMTs are used as inputs to BNN in the paper, not the reconstructed energy and position of events. The samples of neutrino events and three major backgrounds from the Monte-Carlo simulation of a toy detector are generated in the signal region. Compared to the BNN method in Ref.\cite{key-1}, more $^{8}$He/$^{9}$Li background and uncorrelated background in the signal region can be rejected by the BNN method in the paper, but more fast neutron background events in the signal region are unidentified using the BNN method in the paper. The uncorrelated background to signal ratio and the $^{8}$He/$^{9}$Li background to signal ratio are significantly improved using the BNN method in the paper in comparison with the BNN method in Ref.\cite{key-1}. But the fast neutron background to signal ratio in the signal region is a bit larger than the one in Ref.\cite{key-1}.