Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Exploring an image-based $b$-jet tagging method using convolution neural networks

Hangil Jang, Sanghoon Lim

TL;DR

This work tackles jet flavor tagging in heavy-ion collisions by introducing an image-based CNN that encodes track information inside a jet cone as 2D line- and dot-image representations around the primary vertex. A ResNet-18 classifier is trained on simulated jets (generated with $\\sqrt{s}=5$ TeV using PYTHIA8) under ideal and ALICE ITS3-like realistic tracking, producing a discriminant $D_b$ to separate $b$-, $c$-, and light-jets. The line-image representation consistently outperforms the dot-image approach, achieving ROC-AUC values near $0.97$–$0.98$ in ideal conditions and robust performance under detector realism, with $b$-jet efficiency around the 60–80% range at fixed light-jet rejection. This indicates a promising, image-based pathway to improve heavy-flavor jet tagging in heavy-ion experiments, with potential applicability to ALICE ITS3 measurements and deeper insights into jet-medium interactions.

Abstract

Jet flavor tagging, the identification of jets originating from $c$-quarks, $b$-quarks, and other quarks (light quarks and gluons), is a crucial task in high-energy heavy-ion physics, as it enables the investigation of flavor-dependent responses within the hot and dense nuclear medium produced in heavy-ion collisions. Recently, several methods based on deep learning techniques, such as deep neural networks and graph neural networks, have been developed. These deep-learning-based methods demonstrate significantly improved performance compared to traditional methods that rely on track impact parameters and secondary vertices. In the tagging algorithms, various properties of jets and constituent charged particles are used as input parameters. We explore a new method based on images surrounding the primary vertex, utilizing charged particles within the jet cone, which can be measured using a silicon tracking system. For this initial experimental study, we assume the ideal performance of the tracking system. To analyze these images, we employed convolutional neural networks. The image-based flavor tagging method shows an 80-90% $b$-jet tagging efficiency for jets in the transverse momentum range from 20 to 100 GeV/$c$. This approach has the potential to significantly improve the accuracy of jet flavor tagging in high-energy nuclear physics experiments.

Exploring an image-based $b$-jet tagging method using convolution neural networks

TL;DR

This work tackles jet flavor tagging in heavy-ion collisions by introducing an image-based CNN that encodes track information inside a jet cone as 2D line- and dot-image representations around the primary vertex. A ResNet-18 classifier is trained on simulated jets (generated with TeV using PYTHIA8) under ideal and ALICE ITS3-like realistic tracking, producing a discriminant to separate -, -, and light-jets. The line-image representation consistently outperforms the dot-image approach, achieving ROC-AUC values near – in ideal conditions and robust performance under detector realism, with -jet efficiency around the 60–80% range at fixed light-jet rejection. This indicates a promising, image-based pathway to improve heavy-flavor jet tagging in heavy-ion experiments, with potential applicability to ALICE ITS3 measurements and deeper insights into jet-medium interactions.

Abstract

Jet flavor tagging, the identification of jets originating from -quarks, -quarks, and other quarks (light quarks and gluons), is a crucial task in high-energy heavy-ion physics, as it enables the investigation of flavor-dependent responses within the hot and dense nuclear medium produced in heavy-ion collisions. Recently, several methods based on deep learning techniques, such as deep neural networks and graph neural networks, have been developed. These deep-learning-based methods demonstrate significantly improved performance compared to traditional methods that rely on track impact parameters and secondary vertices. In the tagging algorithms, various properties of jets and constituent charged particles are used as input parameters. We explore a new method based on images surrounding the primary vertex, utilizing charged particles within the jet cone, which can be measured using a silicon tracking system. For this initial experimental study, we assume the ideal performance of the tracking system. To analyze these images, we employed convolutional neural networks. The image-based flavor tagging method shows an 80-90% -jet tagging efficiency for jets in the transverse momentum range from 20 to 100 GeV/. This approach has the potential to significantly improve the accuracy of jet flavor tagging in high-energy nuclear physics experiments.

Paper Structure

This paper contains 8 sections, 2 equations, 17 figures, 1 table.

Table of Contents

  1. Introduction
  2. Simulation setup
  3. Neural Network Architecture
  4. Model architecture
  5. Training
  6. Results
  7. Summary
  8. Additional figures

Figures (17)

  • Figure 1: Example 2D images of a light-jet with lines (left) and dots (right) with perfect tracking efficiency and momentum resolution.
  • Figure 2: Example 2D images of a $b$-jet with lines (left) and dots (right) with perfect tracking efficiency and momentum resolution.
  • Figure 3: Example 2D images of a light-jet with lines (left) and dots (right) with realistic tracking efficiency and momentum resolution.
  • Figure 4: Example 2D images of a $b$-jet with lines (left) and dots (right) with realistic tracking efficiency and momentum resolution.
  • Figure 5: The neural network architecture of ResNet-18 with dropout.
  • ...and 12 more figures