[Paper Review] b tagging in ATLAS and CMS

L. Scodellaro|arXiv (Cornell University)|Sep 5, 2017

Particle physics theoretical and experimental studies3 citations

TL;DR

This paper presents advanced b tagging algorithms developed by ATLAS and CMS for identifying jets from bottom quark hadronization during Run2 of the LHC. Utilizing multivariate techniques and deep learning—such as RNNs and deep neural networks—these algorithms significantly improve b jet identification efficiency and light-quark rejection, with performance validated on data and optimized for boosted topologies and future high-luminosity conditions.

ABSTRACT

Many physics signals presently studied at the high energy collision experiments lead to final states with jets originating from heavy flavor quarks. This report reviews the algorithms for heavy flavor jets identification developed by the ATLAS and CMS Collaborations in view of the Run2 data taking period at the Large Hadron Collider. The improvements of the algorithms used in 2015 and 2016 data analyses with respect to previous data taking periods are discussed, as well as the ongoing developments in view of the next years of data taking. The measurements of the performance of the algorithms on data as well as the dedicated techniques for the identification of heavy flavor jets in events with boosted topologies are also presented. Finally, the effectiveness of heavy flavor jet identification in the complex environment expected during the high luminosity LHC phase is discussed.

Motivation & Objective

• Improve b jet identification performance in high-energy proton-proton collisions at the LHC using advanced multivariate and deep learning techniques.
• Address the challenges of high pile-up and boosted topologies in Run2 and future High Luminosity LHC (HL-LHC) phases.
• Enhance the accuracy of b tagging by leveraging improved detector instrumentation, such as the new CMS pixel detector.
• Develop and validate new tagging algorithms for boosted Higgs and top quark decays, where decay products overlap.
• Ensure reliable performance measurements through data-driven scale factors and validation in pure top-quark and QCD-enriched samples.

Proposed method

• Employ multivariate analysis (MVA) to combine multiple discriminants—such as track impact parameters, secondary vertex reconstruction, and soft lepton presence—into a single b tagging discriminant.
• Implement recurrent neural networks (RNNs) in ATLAS to process sequential track variables, enabling better modeling of b jet decay topology.
• Use deep neural networks (e.g., DeepCSV, DeepFlavour) in CMS to process a larger set of input variables and improve discrimination over standard algorithms.
• Apply ghost association to match large R=1 calorimetric jets with small R=0.2 track jets for boosted topology reconstruction.
• Develop dedicated double-b tagging algorithms that exploit angular correlations between subjets in fat jets to enhance H→bb identification.
• Measure b tagging efficiency scale factors using data samples enriched in top quark and semileptonic decays to correct for simulation biases.

Experimental results

Research questions

• RQ1 How do MVA-based b tagging algorithms improve b jet identification efficiency and light-quark rejection compared to Run1 methods?
• RQ2 To what extent can deep learning techniques such as RNNs and deep neural networks enhance b tagging performance in high-pile-up environments?
• RQ3 What is the optimal strategy for identifying boosted Higgs bosons decaying to b quarks using large-cone calorimetric jets and track substructure?
• RQ4 How do new detector components—like the upgraded CMS pixel detector—affect b tagging performance and resolution of secondary vertices?
• RQ5 Can b tagging algorithms maintain high performance under the extreme conditions expected during the High Luminosity LHC phase?

Key findings

• The RNNIP tagger in ATLAS, using only track-by-track variables, outperforms traditional impact parameter-based algorithms in b jet identification.
• CMS's DeepCSV algorithm, trained on data from the new 2017 pixel detector, shows a measurable improvement in performance over its predecessor, with a 4% efficiency gain at 0.1% mis-tagging rate.
• The new DeepFlavour tagger in CMS is projected to achieve a 4% efficiency gain over DeepCSV at a 0.1% mis-tagging probability, based on preliminary results.
• For boosted H→bb decays, requiring two matched track jets to satisfy asymmetric b tagging criteria yields the highest performance, outperforming single-jet or subjet tagging.
• The double-b tagger in CMS, which exploits angular correlations between b quark decay products, outperforms standard fat jet and subjet b tagging in identifying H→bb decays.
• Performance scale factors for b tagging in ATLAS and CMS are measured with high precision using data from t¯t and muon-enriched samples, enabling reliable calibration in physics analyses.