Machine Learning-Based b-Jet Tagging in pp Collisions at s \sqrt{s} = 13 TeV
Hadi Hassan 1, Neelkamal Mallick 1, D.J. Kim 1,2 · 2025
Read Machine Learning-Based b-Jet Tagging in pp Collisions at s \sqrt{s} = 13 TeV by Hadi Hassan 1, Neelkamal Mallick 1, D.J. Kim 1,2 (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2504-18291.txt.
Introduction
Hassan, Mallick, and Kim study how to tag beauty jets in proton collisions in this 2025 paper. A beauty jet is a spray of particles that began as a bottom quark. Tagging such jets tests theory and gives a baseline for heavy ion work. Their data come from proton-proton collisions at an energy of thirteen tera electron volts. Jets were rebuilt from charged particles with a standard clustering method.
The tagger is a machine learning model built on a convolutional neural network. It learns from the jet itself, from its tracks, and from secondary vertices where particles decayed away from the main collision point. The authors report that this model beats older tagging methods in head to head checks. They present the result as a sharper tool for heavy flavor studies. The work shows how image style learning can sort particle sprays that look much alike to simpler rules.
Catalog record
| CREATOR | Hadi Hassan 1, Neelkamal Mallick 1, D.J. Kim 1,2 |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 18 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2504-18291
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Machine Learning-Based b-Jet Tagging in pp Collisions at s \sqrt{s} = 13 TeV Hadi Hassan 1, Neelkamal Mallick 1, D.J. Kim 1,2 Abstract Studying heavy-flavor jets in pp collision is important since they can test pQCD calculations and be used as a reference for heavy-ion collisions. Jets in this analysis are reconstructed from charged particles using the anti- k T k_{\mathrm{T}} algorithm with a resolution parameter R = R= 0.4 and with pseudorapidity | η | < |\eta|< 0.5. Beauty jets are tagged using a machine learning model that uses a convolutional neural network trained on information extracted from the jet, tracks, and secondary vertices. The results show that this model is superior compared to other traditional tagging methods. I Introduction Heavy-Flavor (HF) jets are important for the understanding of quantum chromodynamic (QCD) Politzer:1974fr . They can provide tests for pQCD calculation and have a large mass that acts as a hard scale that suppresses soft-gluon radiations at small angles Dokshitzer:1991fd ; ALICE:2021aqk . They are produced very early in the collision to carry information on the hard collision and information about the high-energy-density medium produced in heavy-ion collisions Zhang:2003wk ; Dokshitzer:2001zm . Identification of jets formed from hadronization of the heavy b-quarks, so-called ”b-jets”, is important for measurements like cross section CMS:2012pgw ; ALICE:2021wct ; CMS:2021pcj , fragmentation, and substructure CMS:2020geg ; CMS:2022btc , and for understanding the quark-gluon plasma (QGP) in heavy-ion collisions ATLAS:2022agz ; CMS:2013qak . Various tagging algorithms have been developed to achieve the identification. The problem with traditional b-jet tagging algorithms, such as impact parameter (IP) or secondary vertex (SV) methods ALICE:2021wct ; CMS:2012feb ; ATL-PHYS-PUB-2017-011 ; ATL-PHYS-PUB-2017-013 that rely on a long lifetime of the b-hadron, is that they rely on cuts posed on single variables. Machine learning (ML) methods can combine several parameters from different levels, like tracks, jets, or secondary vertices, and extract complex multidimensional correlations and patterns among the event characteristics. Previous studies from ATLAS and CMS experiments show that the ML models are superior compared to the traditional methods in terms of performance ATLAS:2023gog ; ATLAS:2022qxm ; app122010574 ; Bols:2020bkb . In this paper, Pythia8 Monte Carlo (MC) generator is used pythia8 ; Bierlich:2022pfr , mainly since it is widely used and carefully validated by the LHC experiments ALICE:2020swj ; CMS:2019csb . Real experimental measurements, such as those performed by ALICE, are affected by finite tracking and pointing resolutions and acceptance, and detector efficiency corrections. Although we present analysis methods that can be optimized for the ALICE environment, we wish to keep the discussion at the methodological level. Hence, we do not perform a complete Monte Carlo (MC) simulation of the ALICE detector ALICE using GEANT4 AGOSTINELLI2003250 , but instead smear the transverse momentum ( p_{\rm T} ) and distance to the closest approach (DCA) distributions to mimic realistic experimental conditions in ALICE. This paper is organized as follows. Sec. II describes the event generation, spectra smearing, jet reconstruction, feature extraction, and machine learning (ML) methodology. Section III presents the performance of the ML algorithm for the tagging of beauty jets and compares it with traditional methods such as the secondary vertex (SV) and impact parameter methods. And finally Sec. IV summarizes the findings of this analysis. II Methodology II.1 Event generation and jet reconstruction Proton-proton (pp) collision events at center-of-mass energy \sqrt{s} = 13 TeV are generated using PYTHIA8 pythia8 ; Bierlich:2022pfr MC event generator with the Monach 2013 tune Skands:2014pea . For a more accurate description of the ALICE detector ALICE the primary vertex is required to be within |z|< 10 cm in the beam direction and only final state (after parton showers and hadronization and decays) charged particles with transverse momentum p_{\rm T} > 0.15 GeV/ c and restricted to ALICE tracker acceptance |\eta|\,< 0.9 are used in this analysis. Then a transverse momentum smearing was applied on these tracks based on the ALICE inner tracking system 2 (ITS2) momentum resolution Abelev:1625842 using the equation: \sigma(p_{\rm T})/p_{\rm T}=\sqrt{a^{2}+b^{2}/p_{\rm T}^{2}+c^{2}.p_{\rm T}^{2}} (1) where a= 0.005, b= 8e-4 GeV/ c , and c= 0.001 (GeV/ c ) -1 are constants that represents the momentum resolution of the ALICE ITS2 detector. In the addition to the p_{\rm T} smearing, a smearing of the distance of the closest approach (DCA) of the particles with a respect to primary vertex has been applied according to the DCA resolution of the ITS2 detector Abelev:1625842 ; Triolo_2025 \sigma : \sigma(DCA_{xy/z})\,(\mu\mathrm{m})=\sqrt{a^{2}+b^{2}/p_{\rm T}^{2}} (2) where a= 7 and 6 \mu m, and b= 23 and 25 \mu mGeV/ c are the constants for the DCA resolution in the transverse plane and z direction, respectively. After selecting the particles and smearing their p_{\rm T} and DCA, they are given to jet clustering algorithm. Jets are reconstructed using the anti- k_{\mathrm{T}} jet clustering algorithm Cacciari:2008gp in the FastJet package Cacciari:2011ma with a resolution parameter R= 0.4, which ensures that most of the initial partonic energy is contained inside the jet. The energies of the tracks are combined using the E-recombination scheme, which considers the mass of the particles with the best strategy. It should be noted here that the real mass of the particles has been used for the jet reconstruction. A jet is classified as a beauty jet (b-jet) if it contains a beauty quark (from the Pythia stack), a charm jet (c-jet) if it contains a charm quark, and a light-flavor jet (lf-jet) if neither quark is present. Jets with beauty quarks coming from gluon splitting are considered as b-jets. II.2 Feature extraction For the machine learning algorithm, several features are extracted from the jet, tracks, and secondary vertices. For jets, the extracted features are the jet transverse momentum p_{\rm T} , pseudo-rapidity \eta , azimuthal angle \varphi , number of track constituents, number of reconstructed secondary vertices (explained later), and the jet mass. For the constituent of the jet, the extracted features are particle p_{\rm T} , particle \eta , the dot product of the momenta of the track and jet, the dot product of the track and the jet divided by the jet momentum, the distance between the track and the jet axis in \eta-\varphi plane calculated as \Delta R=\sqrt{(\eta_{\mathrm{jet}}-\eta_{\mathrm{track}})^{2}-(\varphi_{\mathrm{jet}}-\varphi_{\mathrm{track}})^{2}} . In addition, other features such as the signed DCA in the transverse plane ( xy ), the signed DCA in z direction, and the distance between the track and the closest reconstructed secondary vertex are calculated with the same distance equation mentioned before. The DCA sign is calculated as the sign of the dot product of the jet axis and the DCA vector. Tracks coming from secondary vertices have positive DCA signs due to their long flight distance, while primary tracks coming from the primary vertex could have positive or negative DCA signs. Some of these features extracted from the jet and tracks are very discriminative between the different jet flavors, especially the signed DCA of the track concerning the primary vertex ,as seen on Fig 1 . (a) (b) Figure 1 : The transverse (top) and longitudinal (bottom) DCA distributions for different jet flavors. Prior to particle selection, a maximum DCA (in both xy and z ) of 1 cm is applied to reject tracks originating from strange particle decays. Since the