Transforming jet flavour tagging at ATLAS

Abstract

Jet flavour tagging enables the identification of jets originating from heavy-flavour quarks in proton-proton collisions at the Large Hadron Collider, playing a critical role in its physics programmes. This paper presents GN2, a transformer-based flavour tagging algorithm deployed by the ATLAS Collaboration that represents a different methodology compared to previous approaches. Designed to classify jets based on the flavour of their constituent particles, GN2 processes low-level tracking information in an end-to-end architecture and incorporates physics-informed auxiliary training objectives to enhance both interpretability and performance. Its performance is validated in both simulation and collision data. The measured c-jet (light-jet) rejection in data is improved by a factor of 3.5 (1.8) for a 70% b-jet tagging efficiency, compared to the previous algorithm. GN2 provides substantial benefits for physics analyses involving heavy-flavour jets, such as measurements of Higgs boson pair production and the couplings of bottom and charm quarks to the Higgs boson, and demonstrates the impact of advanced machine learning methods in experimental particle physics.

Fig. 1. Illustration of the GN2 algorithm with jet and track input variables, discriminating between jet flavours by exploiting secondary vertices and other properties stemming from the displaced decays of b-hadrons, in the transverse plane.

The jet features are copied for each track associated with the jet. The combined vectors are then fed into a per-track initialisation network, followed by a transformer encoder and a global representation of the jet. n_jf (n_tf) corresponds to the number of jet (track) features. The pooled jet representation and output track embeddings are provided as inputs to the three task-specific networks. Details of the GN2 architecture are summarised in the ‘Methods’ section.

Fig. 2. b-tagging performance of GN2 and DL1d evaluated in MC simulations.

The c-jet (solid), light-jet (dotted-dashed), and τ-jet (dashed) rejections as a function of the b-jet tagging efficiency for a jets in the $t\overline{t}$ sample with 20 < p_T < 250 GeV and b jets in the $Z^{\prime }$ sample with 250 < pt < 6000 GeV, for both GN2 (light blue) and DL1d (dark orange). The performance of GN2 with respect to DL1d is shown in the bottom panels. The 68% confidence intervals calculated assuming no correlations between the rejections are indicated by the shaded regions, and the uncertainty on each rejection is obtained according to a binomial distribution.

Fig. 3. c-tagging performance of GN2 and DL1d evaluated in MC simulations.

The b-jet (solid), light-jet (dotted-dashed), and τ-jet (dashed) rejections as a function of the c-jet tagging efficiency for a jets in the $t\overline{t}$ sample with 20 < p_T < 250 GeV and b jets in the $Z^{\prime }$ sample with 250 < pt < 6000 GeV, for both GN2 (light blue) and DL1d (dark orange). The performance of GN2 relative to DL1d is shown in the bottom panels. The 68% confidence intervals calculated assuming no correlations between the rejections are indicated by the shaded regions, and the uncertainty on each rejection is obtained according to a binomial distribution.

Fig. 4. b-tagging performance of GN2 and DL1d measured in data and MC simulations.

The a light-jet rejection and b c-jet rejection as a function of the b-jet tagging efficiency for GN2 (light blue) and DL1d (dark orange), directly obtained in simulation (hollowed circle) and rescaled to match those in collision data (solid point). The horizontal error bands correspond to the uncertainties associated with the b-jet tagging efficiency measurement, while the vertical error bands indicate the uncertainties associated with the rejection measurements. A $t\overline{t}$ MC simulation sample with a reconstructed electron or muon is used to derive these results.

Fig. 5. Secondary vertex properties reconstructed using tracks grouped by the GN2 and SV1 algorithms.

The a transverse displacement and the b mass of the secondary vertex obtained by the GN2 (solid) and the SV1 (dotted) algorithms. While the transverse displacement is calculated via a Billoir fit performed on the tracks assigned to the vertex by the respective algorithm, the vertex mass is defined as the invariant mass of the same set of assigned tracks. MC truth (dashed) corresponds to an inclusive reference vertex derived from all tracks associated to simulation-level vertices containing only b-hadron tracks. The last bin in each plot includes overflow.

Fig. 6. The rate of inclusive vertices reconstructed by the GN2 algorithm in light-jets as a function of the jet p_T, without any selections (dotted-dashed) and with the requirement of the vertex containing at least one track with predicted HF origin (solid).

Results from the SV1 algorithm are added as a reference (dotted). The 68% confidence intervals calculated according to a binomial distribution are indicated by the shaded regions.