Performance of heavy-flavour jet identification in Lorentz-boosted topologies in proton-proton collisions at $\sqrt{s}$ = 13 TeV

Basic Information

• Paper ID: 2510.10228
• Title: Performance of heavy-flavour jet identification in Lorentz-boosted topologies in proton-proton collisions at $\sqrt{s}$ = 13 TeV
• Authors: CMS Collaboration
• Classification: physics.ins-det hep-ex
• Publication Date: October 14, 2025
• Journal: Journal of Instrumentation (under submission)
• Paper Link: https://arxiv.org/abs/2510.10228

Abstract

This paper investigates the performance of heavy-flavour hadron jet identification in highly Lorentz-boosted topologies, which is crucial for probing Higgs boson properties and searching for beyond-Standard-Model particles at the LHC. The CMS Collaboration has developed multiple advanced tagging algorithms to identify hadronic jets from the decay of massive particles into $\mathrm{b\overline{b}}$ or $\mathrm{c\overline{c}}$ pairs. The paper primarily demonstrates the performance of these algorithms on simulated events and summarizes novel calibration techniques using proton-proton collision data collected at $\sqrt{s}$ = 13 TeV during the 2016-2018 LHC running period.

Research Background and Motivation

Physics Background

1. Boosted topologies in high-energy physics: At the TeV scale, decay products of heavy particles (such as Higgs bosons and beyond-Standard-Model particles) carry high momentum, causing their decay products to be collimated within a single large-radius jet
2. Importance of heavy-flavour jet tagging: Accurate identification of $\mathrm{b\overline{b}}$ and $\mathrm{c\overline{c}}$ jets is critical for Higgs physics studies and new physics searches
3. Calibration requirements: Discrepancies exist between jet tagging efficiencies in simulated events and actual data, necessitating precise data-driven calibration methods

Research Motivation

1. Precision Standard Model measurements: Precise measurement of Higgs boson decay to heavy-flavour quarks
2. New physics searches: Discovery of new resonant states decaying to heavy-flavour quark pairs
3. Detector performance optimization: Enhancement of the CMS detector's physics object reconstruction performance in boosted topologies

Core Contributions

1. Comprehensive performance assessment: First comprehensive comparison of seven heavy-flavour jet tagging algorithms developed by CMS during Run 2
2. Innovative calibration methods: Development of three independent data-driven calibration approaches:
   • sfBDT method (machine learning-based gluon-splitting jet selection)
   • μ-tagging method (utilizing soft muons within jets)
   • Boosted Z boson method (employing Z→bb decay)
3. Precise scale factor measurements: Combination of multiple measurements via the BLUE method, providing high-precision efficiency correction factors
4. Comprehensive systematic uncertainty assessment: Thorough evaluation of various systematic uncertainty sources and their impacts

Methodology Details

Task Definition

Input: Physical characteristics of large-radius jets (AK8 jets, R=0.8) Output: Jet origin classification probabilities (X→bb, X→cc, QCD, etc.) Objective: Maximize signal efficiency while suppressing QCD multijet background while maintaining mass decorrelation

Tagging Algorithm Architecture

1. ParticleNet-MD

• Architecture: Graph neural network-based particle-level feature processing
• Input: Kinematic and geometric features of particle flow candidates and secondary vertices
• Innovation: Permutation-invariant convolutional operations with local feature extraction in η-φ space
• Output: Mass-decorrelated probability scores

2. DeepDoubleX

• Architecture: Combination of 1D convolutional layers and gated recurrent units
• Feature engineering: Feature selection using layer-wise relevance propagation techniques
• Mass decorrelation: Achieved through reweighting to match signal jet mass distribution to QCD background

3. DeepAK8-MD

• Architecture: Multi-class classifier based on 1D residual convolutional layers
• Adversarial training: Mass decorrelation implemented using a mass prediction network as a penalty term in the loss function

4. Double-b Tagger

• Architecture: Boosted decision tree (BDT)-based approach
• Features: High-level track and secondary vertex construction variables

Calibration Methods

1. sfBDT Method

Core concept: Use BDT to select gluon-splitting bb/cc jets similar to signal jets as proxies

Key innovations:
- Definition of hadron-level N-subjettiness variable τ^h_31 to distinguish signal from background
- Automated sfBDT selection threshold determination procedure
- 81 different selection combinations for systematic uncertainty assessment

2. μ-Tagging Method

Physical principle: Semi-leptonic decay modes of b(c) hadrons produce soft muons
Selection criteria:
- Presence of soft muons with pT > 5 GeV within the jet
- τ21 < 0.3 (selecting double-prong jet structure)
- Relative isolation Irel > 0.15

3. Boosted Z Boson Method

Signal extraction: Extract Z→bb signal peak from QCD multijet background
Fitting strategy:
- 2D fit (mPNet, pT)
- QCD background modeled with polynomial function
- Simultaneous fit of regions passing and failing tagger selection

Experimental Setup

Datasets

• Experimental data: Proton-proton collision data collected by CMS in 2016-2018
  • 2016 pre-VFP: 19.5 fb⁻¹
  • 2016 post-VFP: 16.8 fb⁻¹
  • 2017: 41.5 fb⁻¹
  • 2018: 59.8 fb⁻¹
• Simulated samples:
  • QCD multijet processes (MADGRAPH5 aMC@NLO)
  • V+jets processes (Z+jets, W+jets)
  • Higgs boson production (HJ-MINLO + PYTHIA)

Evaluation Metrics

• Signal efficiency: Fraction of correctly tagged X→bb(cc) jets
• Background rejection: Fraction of incorrectly tagged QCD jets
• Scale factors (SF): Ratio of data to simulation efficiency SF = ε_data/ε_sim
• ROC curves: Trade-off relationship between signal efficiency and background efficiency

Working Point Definition

Three working points defined for each tagging algorithm:

• High purity (HP): 40%(bb)/15%(cc) signal efficiency
• Medium purity (MP): 60%(bb)/30%(cc) signal efficiency
• Low purity (LP): 80%(bb)/50%(cc) signal efficiency

Experimental Results

Algorithm Performance Comparison

Scale Factor Measurement Results

ParticleNet-MD X→bb (2018 Data)

Inter-method Consistency

Results from the three calibration methods remain consistent within uncertainty ranges:

• sfBDT method: Generally yields higher SF values
• μ-tagging method: Intermediate SF values with larger uncertainties
• Boosted Z boson method: Limited by statistics, largest uncertainties

Systematic Uncertainty Decomposition

Primary uncertainty sources (using ParticleNet-MD HP WP as example):

1. Statistical uncertainty: ~6%
2. sfBDT selection dependence: ~5%
3. Reweighting scheme effects: ~9%
4. Theoretical uncertainty (ISR/FSR): ~1-4%

Related Work

Traditional Methods

• BDT with high-level variables: Using manually engineered jet shape variables
• Simple b-tagging: Based on secondary vertex and track information

Deep Learning Method Evolution

1. DeepCSV/DeepJet: Deep learning tagging for AK4 jets
2. CNN approaches: Image-based jet processing
3. Graph neural networks: Direct particle-level information processing
4. Transformer architectures: Attention mechanisms in jet tagging

Calibration Method Development

• Early methods: Based on simple kinematic selections
• Template fitting: Signal extraction using invariant mass spectra
• ML-assisted approaches: Using ML methods to improve proxy jet selection

Conclusions and Discussion

Main Conclusions

1. ParticleNet-MD shows optimal performance: Achieves best performance in both X→bb and X→cc tagging tasks
2. Neural networks outperform traditional methods: Deep learning methods significantly surpass traditional BDT-based approaches
3. Calibration methods are effective: Three independent methods provide consistent scale factor measurements
4. Mass decorrelation successfully implemented: All modern algorithms successfully achieve decorrelation from jet mass

Limitations

1. Statistical precision constraints: Particularly in high-pT regions and high-purity working points
2. Systematic uncertainties: Primarily from model dependence in proxy jet selection
3. Applicability scope: Calibration results mainly applicable to similar boosted topologies
4. Computational complexity: Deep learning methods incur higher computational costs

Future Directions

1. Run 3 data analysis: Improved measurement precision using larger statistical samples
2. New architecture exploration: Novel neural network architectures such as Transformers
3. End-to-end optimization: Full-chain optimization from detector signals to physics analysis
4. Real-time applications: Implementation of advanced jet tagging in trigger systems

In-Depth Evaluation

Strengths

1. Comprehensive scope: First comprehensive comparison of all major CMS heavy-flavour jet tagging algorithms
2. Methodological innovation: Three independent calibration methods provide mutual verification, enhancing result reliability
3. Advanced technology: Represents the state-of-the-art in jet tagging technology
4. High practical value: Provides important calibration tools for CMS physics analyses
5. Complete uncertainty assessment: Systematic evaluation of various uncertainty sources

Weaknesses

1. Limited theoretical understanding: Lacks deep physical insight into why certain methods perform better
2. Insufficient computational efficiency discussion: Limited discussion of computational cost trade-offs among algorithms
3. Limited generalization assessment: Insufficient evaluation of algorithm generalization across different physics processes
4. Statistical limitations: Some measurements constrained by statistical precision

Impact

1. Academic influence: Establishes new standards for jet tagging technology in high-energy physics experiments
2. Practical value: Directly serves Higgs physics and new physics searches
3. Technology transfer: Methods generalizable to other experiments and object identification tasks
4. Industrial application potential: Deep learning techniques applicable to other pattern recognition problems

Applicable Scenarios

1. Higgs physics research: Precision measurements of H→bb, H→cc decay channels
2. New physics searches: Discovery of new resonant states decaying to heavy-flavour quarks
3. Precision measurements: Analyses requiring high-precision heavy-flavour jet identification
4. Methodological research: Benchmark testing and comparison of jet tagging algorithms

Technical Innovation Highlights

sfBDT Method Innovation

• Hadron-level τ^h_31 variable: First use of N-subjettiness based on first-generation hadrons to distinguish signal from background
• Automated threshold selection: Development of algorithm for automatic optimal sfBDT selection determination
• Multiple selection strategies: Quantification of selection-dependent systematic uncertainties through 81 selection combinations

Combined Measurement Technique

• BLUE method extension: Extension of Best Linear Unbiased Estimate method to simultaneous fitting across multiple pT intervals
• Correlation handling: Proper treatment of systematic uncertainty correlations among different methods
• Cross-validation: Three independent methods provide strong mutual validation

References

The paper cites 72 important references, covering:

• CMS detector technical literature
• Historical development of jet tagging algorithms
• Deep learning applications in high-energy physics
• Statistical methods and uncertainty treatment
• Related physics analysis results

Overall Assessment: This is a high-quality experimental physics paper representing the state-of-the-art in jet tagging technology in particle physics experiments. The paper not only provides important technical tools but also establishes a solid foundation for future algorithm development and physics analyses. Its methodological innovations and systematic performance evaluation hold significant value for the entire high-energy physics community.