Skip to main content
. .

ATLAS Flavour Tagging

The identification of jets originating from bottom, charm, and other quarks is essential for many measurements and searches at the ATLAS experiment at CERN. This technique, known as flavour tagging, is particularly important for studies of the Higgs boson, top quarks, and rare processes that may reveal physics beyond the Standard Model.

At the University of Siegen, we develop advanced algorithms for jet flavour identification, secondary-vertex reconstruction, and their calibration using collision data. Our research combines detector reconstruction with modern machine-learning methods, particularly transformer networks. We contribute to the development of ATLAS flavour-tagging algorithms and their application in precision measurements and searches for new physics.

How does flavour tagging work?

Bottom and charm quarks form hadrons that have characteristic properties distinguishing them from particles produced by light quarks and gluons. In particular, hadrons containing heavy quarks have relatively long lifetimes. B-hadrons, for example, have typical lifetimes of approximately 1.5 ps and can travel several millimetres before decaying.

Their decay products therefore frequently originate from a displaced secondary vertex rather than directly from the primary proton-proton collision vertex. The tracks of these particles have characteristic impact parameters that can be measured with the high-resolution silicon tracking detectors of ATLAS. Additional information is provided by the masses, multiplicities, and kinematic properties of the decay products.

Flavour-tagging algorithms combine these observables to distinguish jets containing b-hadrons or c-hadrons from jets originating from lighter partons. A major challenge is to distinguish genuine heavy-flavour decays from secondary vertices produced by interactions with detector material, photon conversions, decays of long-lived light hadrons, and additional proton-proton collisions (pile-up).

Transformer-based flavour tagging: GN2 and GN3

Modern machine-learning techniques have fundamentally transformed flavour identification at ATLAS. Earlier algorithms combined dedicated low-level reconstruction tools with multivariate classifiers. Newer approaches process information from individual tracks and other jet constituents directly, learning the correlations that characterize heavy-flavour decays.

The transformer-based GN2 algorithm uses the properties of tracks associated with a jet to identify its flavour. In addition to jet classification, the network is trained to predict the origins of individual tracks and whether pairs of tracks originate from a common decay vertex. These auxiliary tasks encourage the network to learn physically meaningful representations of heavy-hadron decay structures.

GN2 substantially improves flavour identification compared with previous ATLAS algorithms. At a fixed b-tagging efficiency, the rejection of charm and light-flavour jets can be increased by factors of several, depending on the operating point and jet kinematics. The development and performance of GN2 were documented in a publication in Nature Communications in 2026.

The next generation, GN3, extends this approach by combining multiple types of detector information. In addition to charged-particle tracks, the algorithm incorporates particle-flow objects and soft leptons, allowing it to exploit neutral and semileptonic components of jet substructure. Improvements in track association and additional training tasks further enhance its performance.

GN3 also extends flavour identification beyond bottom and charm jets by providing separate classification outputs for strange-quark jets, up- and down-quark jets, and gluon jets. This opens new opportunities for identifying the partonic origin of jets and studying the flavour structure of Standard Model processes.

Our group contributes to the development and evaluation of modern flavour-tagging algorithms, including GN3 and its extended flavour-identification capabilities. Members of our group presented GN3 results at ML4Jets 2025 and at the DPG Spring Meeting in 2026.

Machine-learning software development

The development of modern flavour-tagging algorithms requires flexible software for handling complex detector information, constructing neural networks, and training models on large simulation datasets. Our group contributes to Salt, an open-source framework for multimodal, multitask machine learning in high-energy physics.

Salt provides tools for training transformer-based models that can combine different detector inputs and address several prediction tasks simultaneously. It supports the development of algorithms such as GN2 and GN3 and facilitates their application to other particle identification and reconstruction problems.

The framework was described in a publication in the Journal of Open Source Software in 2025, with a contribution from our group. This work forms part of our broader effort to develop modern, reusable machine-learning tools for experimental particle physics.

Secondary-vertex reconstruction

The reconstruction of displaced decay vertices has been a long-standing research focus of our group. We have contributed to the development of the Secondary Vertex Finder (SVF), which reconstructs displaced vertices inside jets, and the Track Classification Tool (TCT), which estimates whether individual tracks originate from heavy-hadron decays, fragmentation, or other sources.

These approaches provided important building blocks for flavour-tagging algorithms, including the development of the Classified Track Tagger (CTT), which aimed to improve the identification of b-jets at high transverse momentum.

More recently, our research has focused on the reconstruction of low-momentum b-hadrons and secondary vertices that cannot necessarily be associated with a reconstructed jet. Such soft secondary vertices are particularly relevant for searches for new particles that produce relatively soft b-hadrons in their decay chains.

Our group contributed to the development and performance studies of soft secondary-vertex reconstruction tools, including their calibration using ATLAS collision data. In 2024, ATLAS published a measurement of the tagging efficiency and mistag rate of a dedicated soft secondary-vertex tagger, using the full Run-2 dataset. The results provide correction factors for simulated events and are relevant for searches involving low-momentum heavy-flavour particles.

Our subsequent work investigated the performance of low-energy secondary-vertex reconstruction, including the influence of track reconstruction and background processes. These studies were part of a doctoral research project completed in Siegen in 2025.

Material interactions and vertex reconstruction

A particular challenge for flavour tagging is the reconstruction of secondary vertices that do not originate from heavy-hadron decays. Charged particles may interact with the material of the tracking detector, producing additional tracks and displaced vertices that can resemble the decay signatures of b- or c-hadrons.

Our group studies the influence of such material interactions on jet flavour identification. These investigations include the reconstruction and classification of secondary vertices, their geometrical distribution within the tracking detector, and the differences between collision data and simulation.

The work was a central topic of a master's thesis completed in 2025 and contributes to a better understanding of backgrounds and systematic uncertainties in flavour tagging.

We also investigate modern algorithms for reconstructing multiple secondary vertices, including approaches based on the ACTS tracking toolkit. Studies using the Open Data Detector (ODD) provide a flexible environment for developing and evaluating vertex-reconstruction methods relevant to future high-luminosity experiments.

Charm tagging and calibration

The identification of charm jets is particularly challenging. Compared with B-hadrons, charm hadrons generally have shorter lifetimes and lower masses, making their displaced decay signatures less pronounced. A successful charm-tagging algorithm must distinguish c-jets both from b-jets and from light-flavour jets.

Our group works on the optimisation of charm-jet identification, including the selection and combination of flavour-tagging discriminants. These developments are especially important for physics measurements that require the separation of charm and bottom quarks.

In addition to improving the algorithms themselves, their performance must be measured using collision data. Differences between data and simulation are accounted for through calibration correction factors and associated systematic uncertainties.

Our recent activities include studies of charm-tagging optimisation presented in 2025 and charm-tagging calibration studies presented in 2026. These efforts are closely connected to recent ATLAS measurements of b-tagging efficiencies and c-jet mistagging rates using Run-3 data at a centre-of-mass energy of 13.6 TeV.

Simulation and robustness of flavour tagging

Modern flavour-tagging algorithms are trained primarily using simulated collision events. Their performance therefore depends on the accuracy with which the simulation describes particle production, detector responses, track reconstruction, and the properties of jets containing heavy-flavour hadrons.

Our group investigates how changes in detector simulation and reconstruction affect the performance of machine-learning flavour taggers. Current studies include the use of fast detector simulation for jet flavour identification and the robustness of modern algorithms against changing experimental conditions.

A central objective is to understand which detector properties and modelling assumptions have the largest impact on classification performance. This information can guide improvements to the simulation, identify important systematic uncertainties, and help develop algorithms that remain reliable under different operating conditions.

Heavy-flavour jet energy calibration

Precise measurements also require an accurate determination of the energy and momentum of jets containing heavy-flavour hadrons. These jets have characteristic fragmentation patterns and may contain neutrinos from semileptonic heavy-hadron decays, which complicate their energy calibration.

Transformer-based regression methods can combine information from individual jet constituents to improve the reconstructed energy and mass of b-jets. ATLAS demonstrated this approach in 2024, achieving significant improvements in jet energy and mass resolution compared with conventional calibration methods.

Although energy calibration and flavour identification address different reconstruction tasks, they exploit related machine-learning approaches and can benefit from a common understanding of jet substructure and detector response.

Applications in top-quark and Higgs-boson physics

Flavour tagging is essential for many physics analyses performed by our group. In top-quark pair (tt̄) production, b-jet identification helps reconstruct top-quark decays and suppress background processes.

A particularly important application is the measurement of the CKM matrix element |Vcb| using the rare decay W → cb̄ in tt̄ events. In 2026, ATLAS published the first measurement of |Vcb| using this method, with a major contribution from our group. The analysis relies on the simultaneous identification of bottom and charm jets.

Improved charm tagging is also central to searches for Higgs-boson decays into charm-quark pairs, H → cc̄, particularly in associated tt̄H production. Such analyses aim to constrain the Higgs coupling to second-generation quarks and require high rejection of backgrounds containing b-jets or light-flavour jets.

Further applications include rare top-quark decays, measurements of Higgs-boson interactions, and searches for new particles. More information can be found on our ATLAS Top-Quark Physics and ATLAS Higgs-Physics pages.

Outlook: flavour tagging at the HL-LHC

The High-Luminosity LHC will present new challenges for flavour tagging. The substantially increased number of simultaneous proton-proton collisions will lead to higher track densities and more complex event topologies. At the same time, the much larger datasets will make precise heavy-flavour identification increasingly important for rare-process measurements.

The upgraded ATLAS Inner Tracker (ITk) will provide improved tracking capabilities under these demanding conditions. Our research on transformer-based flavour tagging, secondary-vertex reconstruction, simulation, and calibration contributes to preparing the ATLAS analysis programme for the HL-LHC era.

Selected publications and research outputs

Recent contributions from Siegen

Selected research activities presented by members of our group at conferences and workshops:

  • 2026: GN3 and flavour tagging beyond heavy quarks; charm-tagging calibration; fast detector simulation for jet flavour identification; improved secondary-vertex reconstruction using ACTS.
  • 2025: GN3 developments and performance; charm-tagging optimisation; material interactions in flavour tagging; soft secondary-vertex reconstruction and calibration.
  • 2024: Soft secondary-vertex reconstruction; calibration of low-momentum heavy-flavour signatures; transformer-based heavy-flavour jet calibration.

See also our DPG Spring Meeting contributions and conference presentations.