When high-energy colliders smash protons together, the resulting quarks and gluons cannot exist in isolation due to the laws of color confinement. Instead, they instantly fragment into collimated sprays of energetic hadrons known as jets. To test the Standard Model or search for signs of new physics—such as Minimal Flavor Violation as explored by Yuval Grossman et al. in Probing minimal flavor violation at the CERN LHC—physicists must determine the exact "flavor" of the initial parton that triggered the jet. This process is known as flavor tagging.
The Physics of Flavor Signatures
Heavy quarks—specifically bottom ($b$) and charm ($c$) quarks—leave distinct experimental signatures because they form hadrons with relatively long lifetimes. These heavy hadrons travel a few millimeters at near-light speed before decaying, creating a "secondary vertex" that is spatially displaced from the primary collision point.
As noted by C. Aidala et al. in Measurement of charm and bottom production from semileptonic hadron decays..., measuring these displaced tracks allows physicists to statistically separate heavy-flavor contributions. However, when particles are highly "boosted" (traveling with extreme transverse momentum), the decay products become highly collimated and dense. As Yuval Grossman et al. point out, this high-momentum regime introduces intense Quantum Chromodynamics (QCD) radiation and complex combinatoric backgrounds, making traditional vertex isolation exceptionally difficult.
The Intermediate Flavor Gap
While $b$-tagging algorithms have achieved high precision, identifying charm ($c$) quarks presents a unique bottleneck. In Heavy-flavour jet identification at the CMS experiment for Run 2, Mauro Verzetti explains that charm quarks are physically "sandwiched" between light flavors and bottom quarks:
- Light quarks ($u, d, s$) and gluons: Virtually zero lifetime, lower mass.
- Charm quarks ($c$): Intermediate lifetime ($\sim 0.1–0.4\text{ ps}$), intermediate mass ($\sim 1.3\text{ GeV}$).
- Bottom quarks ($b$): Longer lifetime ($\sim 1.5\text{ ps}$), higher mass ($\sim 4.2\text{ GeV}$).
Because charm properties lie directly in the middle, standard classification boundaries easily misidentify charm jets as either light jets or $b$-jets. Andrzej Novak highlights in Heavy flavour jet identification with the CMS experiment in Run 2 that this efficiency gap has long hindered searches for rare second-generation fermion interactions, such as the Higgs boson decaying into charm quarks ($H \to c\bar{c}$).
Moving to Multi-Class Deep Learning
To resolve the intermediate flavor gap, modern algorithms have transitioned from simple cut-based selections on hand-crafted variables to deep learning architectures. Instead of evaluating isolated parameters like impact parameter significance, deep models process the entire jet topology.
As detailed by Mauro Verzetti in Reconstruction and identification of hadronic objects with CMS, architectures like Convolutional Neural Networks (CNNs) and Recurrent Neural Networks (LSTMs) treat jet energy deposits as images or sequences of particle tracks. By learning the collective, non-linear correlations of the jet's radiation pattern, these multi-class classifiers can simultaneously distinguish $b$-jets, $c$-jets, and light-flavor/gluon jets, providing the high-purity discrimination required for boosted regimes (Andrzej Novak, Inclusive search for a boosted Higgs boson and observation of the Z boson decaying to charm quarks with the CMS experiment).
Data-Driven Disentanglement via Jet Topics
A fundamental challenge of training deep neural networks for flavor tagging is their reliance on Monte Carlo simulations. Because "quark jets" and "gluon jets" cannot be defined from first principles in a gauge-invariant manner at the observable level, simulations introduce systematic uncertainties.
To bypass this limitation, Eric Metodiev et al. proposed a data-driven framework called "Jet Topics" in Jet Topics: Disentangling Quarks and Gluons at Colliders. Borrowing topic modeling techniques from natural language processing—where documents are treated as mixtures of latent topics—the authors treat experimental jet samples as mixtures of underlying jet flavors.
$$\text{Sample}_A = f_q^A \cdot P(\text{quark}) + f_g^A \cdot P(\text{gluon})$$
$$\text{Sample}_B = f_q^B \cdot P(\text{quark}) + f_g^B \cdot P(\text{gluon})$$
By measuring different physical mixtures (for instance, jets produced alongside a $Z$ boson versus jets from pure dijet events), the algorithm mathematically extracts the pure, independent distributions ($P$) of quark and gluon properties directly from operational data. This eliminates the need for simulation labels and has been extended by Jasmine Brewer et al. in Data-driven quark- and gluon-jet modification in heavy-ion collisions to isolate how distinct jet flavors interact with complex environments like the quark-gluon plasma.
Enforcing Physical Symmetries
While deep neural networks offer high accuracy, they are often computationally expensive "black boxes." A recent alternative focuses on embedding physical symmetries directly into the mathematical architecture of the tagger.
In Boost invariant polynomials for efficient jet tagging, Jose M Munoz et al. introduced Boost Invariant Polynomials (BIPs). Because the underlying physics of a jet is invariant under Lorentz boosts along the beam axis, the authors structured the input features as polynomials that automatically respect this symmetry. By hardcoding relativity into the feature space, BIPs eliminate the need for a network to learn coordinate transformations from scratch, matching the accuracy of deep learning models while drastically reducing training and inference times.
Key Foundational Work
Probing minimal flavor violation at the CERN LHC (2007) — contributed 2 concept elements to this direction.
Measurement of time-dependent CP asymmetry in B0→cc K(*)0 decays (2009) — contributed 1 concept elements to this direction.