Recovery of merged πo,s from ECL images of the Belle II detector using CNN

Abstract

This study aims to utilize a Convolutional Neural Network (CNN) to retrieve merged π 0 mesons lost in the Belle II experiment, where they appear as individual photons. The issue arises when dealing with high momentum π 0 mesons, i.e., beyond 2 GeV in the Belle II experiment, as the shower produced by both the π 0 meson and gamma appear indistinguish- able at the Electromagnetic Calorimeter (ECL) detector. Currently, reconstruction software is utilized to match photon pairs created by the π 0 → γγ decay; however, the efficiency of this process can be affected by the γ produced by the rest of the events (ROEs), which mimic the signal. One of the most challenging tasks in particle physics research is accu- rately identifying and reconstructing subatomic particles. By the nature of the problem and its importance, accurate reconstruction of π 0 mesons is crucial for identifying various B/D

meson decays, including rare decays like D 0 → γγ 8.5 × 10 −7 , D 0 → ρ 0 γ 10 −5 , and

D 0 → φ γ 10 −5 . These rare decays have dominant background arising from decays like

D 0 → K s π 0 1.24 × 10 −2 , D 0 → π 0 π 0 8.26 × 10 −4 , and D 0 → φ π 0 1.17 × 10 −3 . The Convolutional Neural Networks performed reasonably well on a test dataset, which is identical to real scenarios, achieving an area under the curve (AUC) of 0.86 for the Precision-Recall curve. These results demonstrate the potential of machine learning (ML) algorithms and highlight areas for improvement in the current work to enhance the effi- ciency of identifying π 0 particles with energy deposits in the ECL. The findings suggest that the ‘raw’ ECL images contain much more information than currently used expert- engineered features.