Performance and efficiency of a transformer-based quark/gluon jet tagger in the ATLAS experiment

TL;DR

This work introduces DeParT, a transformer-based quark/gluon jet tagger for ATLAS jets across an extended phase space up to $|\eta|<4.5$ with $p_T>20$ GeV. The model processes jet constituents (PFOs and TopoTowers) and uses a training scheme that encompasses multiple $p_T$-and-eta bands to ensure uniform coverage. Jet distributions in data are extracted using two approaches: a conventional MC-based matrix method and a novel jet topics method that reduces reliance on MC modelling; the jet topics approach generally yields smaller uncertainties, enhancing the tagger’s calibration and applicability to precision SM and new-physics analyses. Across Run-2 and Run-3 data, DeParT demonstrates improved discrimination compared to a FC baseline, with consistent results between data and MC and robust performance in forward regions where TopoTowers provide an advantage. Overall, the study delivers a robust q/g tagging framework with reduced systematic uncertainties, supporting high-precision jet physics and searches at the LHC.

Abstract

A deep-learning approach based on the transformer architecture is developed to distinguish between jets originating from quarks and gluons. The algorithm operates on jets with transverse momentum $p_{\text{T}} > 20$ and pseudorapidity $|η| < 4.5$ and takes as input several properties derived from the jet constituents, using information from the ATLAS detector's tracker and calorimeter. The algorithm's performance is evaluated by analyzing dijet data events from proton-proton collisions at $\sqrt{s} = 13$ and $13.6$ TeV during Run 2 and Run 3 of the Large Hadron Collider. Two methods are used to obtain distributions from quark- or gluon-initiated jets in data: a matrix method fully based on Monte Carlo simulation and a new approach named `jet topics' which has less dependence on the modelling of the physics process under study. The quark and gluon identification efficiencies measured in data for the 50% quark-identification-efficiency working point vary from the simulated ones for quark-initiated (gluon-initiated) jets by factors of 0.88-1.30 (0.61-1.05) with uncertainties of 10%-70% (10%-95%). The uncertainties estimated with the jet topics method are smaller than those estimated with the matrix method, with up to 20% less systematic uncertainty in some phase-space regions. The advances in jet identification reported here provide a robust tool for precision Standard Model measurements and searches for new physics at the LHC.

Figures (12)

• Figure 1: The gluon jet rejection at 50% quark jet efficiency as a function of for jets with \ref{['fig:fc_vs_depart_a']}$|\eta| = 0.0{-}1.2$ and \ref{['fig:fc_vs_depart_b']}$|\eta| = 3.2{-}4.5$ for the FC and DeParT models. The bottom panels show the FC-to-DeParT ratios. The error bars represent the statistical uncertainties.
• Figure 2: The DeParT score distributions for quark and gluon jets in MC events compared with \ref{['fig:score_mc_data_run2_ctrl']} Run-2 and \ref{['fig:score_mc_data_run3_ctrl']} Run-3 jets with $0.0 < |\eta| < 1.2$ and $800 < \pt < 1100\ \text{Ge V}\xspace$, and \ref{['fig:score_mc_data_run2_fwd']} Run-2 and \ref{['fig:score_mc_data_run3_fwd']} Run-3 jets with $3.2 < |\eta| < 4.5$ and $110 < \pt < 160\ \text{Ge V}\xspace$. The error bars represent the statistical uncertainties of data and the shaded bands represent the statistical uncertainties of MC events. The bottom panels show the data-to-MC ratios.
• Figure 3: Fractions of quark jets in the forward ($f_\text{F}^q$) and central ($f_\text{C}^q$) samples as estimated from MC simulation for jets with \ref{['fig:fractions_ctrl']}$0.0 < |\eta| < 1.2$ and \ref{['fig:fractions_fwd']}$3.2 < |\eta| < 4.5$. The error bands represent the statistical uncertainties.
• Figure 4: The \ref{['fig:extraction_g']} gluon and \ref{['fig:extraction_q']} quark DeParT score distributions extracted from Run-2 data with the matrix method (solid red), jet topics without MC correction (dash dotted green), and jet topics with MC correction (dashed blue). The bottom panels show ratios of jet topics to matrix method results. The error bands represent statistical uncertainties.
• Figure 5: The MC efficiencies for truth-labelled \ref{['fig:eff_quark_run2_ctrl']} quark jets with $0.0 < |\eta| < 1.2$, \ref{['fig:eff_quark_run2_fwd']} quark jets with $3.2 < |\eta| < 4.5$, \ref{['fig:eff_gluon_run2_ctrl']} gluon jets with $0.0 < |\eta| < 1.2$, and \ref{['fig:eff_gluon_run2_fwd']} gluon jets with $3.2 < |\eta| < 4.5$, compared with the efficiencies measured for Run-2 data with the jet topics method. Gluon efficiencies are reported as $1-\varepsilon_g$ to ease their comparison with the gluon SF measurements. The 50% WP corresponds to a score threshold defined by a fixed 0.5 quark efficiency in the Pythia sample in each and bin. The deviations of the truth quark-jet efficiencies from 0.5 are due to the finite binning of the DeParT score used in the efficiency calculation. Only the statistical uncertainties are shown.