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

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 $|\eta| < 4.5$ 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.