Application of CNN to a fine segmented scintillator detector for a single particle and neutrino-nucleon event

TL;DR

This study demonstrates that applying CNNs to high-resolution scintillator detector data significantly improves particle momentum resolution and classification accuracy in neutrino experiments.

Contribution

The paper introduces CNN-based methods for event classification, momentum regression, and hit segmentation in a 1 cm subdivided scintillator detector for neutrino research.

Findings

CNN improves momentum resolution by a factor of 2.

Classification accuracy is approximately 94% for single particles.

Feasibility of cross-section analysis with CNN is shown.

Abstract

This paper presents studies on application of convolutional neural network (CNN) to GEANT4 optical simulation data generated with a scintillator detector subdivided into 1 cubic cm, which is designed for the long-baseline neutrino experiment. Classification of interaction, regression of momentum, and segmentation of hits are demonstrated for single particle and neutrino-nucleon interaction events with well established CNN architectures by feeding reconstructed 2D projection images. In the study it is shown that the application of CNN to the 1 cm subdivided scintillator detector can provide a factor about 2 better momentum resolution compared to a standard method, as well as a classification capability of about 94% for the single particle and 70% for the neutrino-nucleon interaction events. Cross-section analyses with CNN is also shown to be feasible.