Abstract
Identification of jets originating from $b$ or $c$ quarks is important for a wide variety of standard model physics (e.g. final states involving top quarks and Higgs bosons) as well as searches for physics beyond the standard model. Due to the increased center-of-mass energy during the LHC Run 2, final states with boosted $b$ quark jets became increasingly important. The CMS Collaboration improved on the existing $b$ jet identification algorithms, and developed the first charm jet tagger. To address boosted topologies a boosted double-$b$ tagger has been developed as well. A review of these new developments is presented, together with their performance on proton-proton collision data recorded by the CMS detector at a centre-of-mass energy of $\sqrt{s} = 13$ TeV during 2015 and 2016.
Introduction
In high-energy particle collisions at the Large Hadron Collider (LHC), quarks and gluons produced in the interaction do not appear as isolated particles. Instead, they undergo a process called hadronization, forming collimated sprays of particles known as "jets." Identifying the original "flavour" of the quark that initiated a jet—specifically distinguishing $b$ (bottom) and $c$ (charm) quarks from light-flavour quarks ($u, d, s$) and gluons—is a cornerstone of modern particle physics. This process, known as jet tagging, is essential for studying the Higgs boson, top quarks, and searching for new physics phenomena.
LHC Run 2 presented new challenges. With the center-of-mass energy reaching 13 TeV, particles are frequently produced with extremely high momentum. In this "boosted" regime, the decay products of heavy particles (like a Higgs boson decaying into two $b$ quarks) can merge into a single, wide "fat jet," rendering traditional resolved-jet algorithms ineffective.
Heavy-Flavour Tagging Algorithms
To meet these challenges, the CMS Collaboration developed an upgraded suite of machine-learning-based identification tools.
CSVv2 and cMVAv2
The Combined Secondary Vertex v2 (CSVv2) algorithm is an evolution of Run 1 techniques. It utilizes a Multi-Layer Perceptron (MLP) to integrate information from displaced tracks and reconstructed secondary vertices. Building upon this, the Combined MVA v2 (cMVAv2) acts as a "tagger of taggers." It employs a Boosted Decision Tree (BDT) to aggregate outputs from CSVv2, Jet Probability (JP) algorithms, and soft lepton taggers, exploiting subtle correlations to maximize separation power.
The Charm Tagger
Charm quarks are particularly difficult to identify because their mass and lifetime are intermediate between light quarks and $b$ quarks. CMS introduced the first dedicated charm tagger, which utilizes two correlated BDTs: CvsL (separating $c$-jets from light-flavour jets) and CvsB (separating $c$-jets from $b$-jets). Because $c$-jets occupy an intermediate state, they are best analyzed in a 2D discriminator space.
Boosted Double-b Tagger
For high-energy "fat jets" (reconstructed using the AK8 algorithm), CMS developed the Boosted Double-$b$ Tagger. Unlike standard taggers that look for a single $b$-hadron, this algorithm is specifically trained to identify the presence of two $b$-hadrons within the same wide jet, a characteristic signature of boosted $H \to b\bar{b}$ decays.
Performance and Results
The performance of these algorithms was validated using proton-proton collision data at $\sqrt{s} = 13$ TeV. Initial evaluations were performed on $t\bar{t}$ and QCD multi-jet simulations, followed by validation with up to 7.7 $fb^{-1}$ of 2016 data.
As shown in the performance curves, the transition to complex Multi-Variate Analysis (MVA) ensembles provided a significant leap in efficiency. The cMVAv2 algorithm consistently outperforms its predecessors, offering lower misidentification probabilities for a given $b$-jet efficiency.
Key findings from the Run 2 performance study include:
- Ensemble Learning: The use of BDTs and MLPs allows the taggers to capture complex correlations that simple track-based methods miss.
- Data/Simulation Agreement: The "Scale Factors" (the ratio of efficiency in real data versus simulation) remained close to 1.0, indicating that the complex MVA models are well-modeled and reliable for physics measurements.
- Background Rejection: The dedicated double-$b$ tagger achieved a marked reduction in mistagging rates from QCD background compared to previous subjet-based approaches.
Conclusion and Future Outlook
While the Run 2 toolkit represents a major advancement, heavy-flavour identification remains an active area of research. The complexity of calibrating charm taggers—requiring new methods like $W+c$ and semileptonic $t\bar{t}$ analysis—highlights the ongoing challenge of intermediate-flavour tagging. Furthermore, as the LHC moves toward even higher luminosities and energies, further optimization of taggers for extremely high-$p_T$ regimes will be essential for the next generation of physics discoveries.
Summary updated