Graph Neural Network Jet Flavour Tagging with the ATLAS Detector

Flavour tagging, the identification of jets originating from $b$- and $c$-quarks, is a critical component of the physics programme of the ATLAS experiment at the Large Hadron Collider. Current flavour tagging algorithms rely on the outputs of several low-level algorithms, which reconstruct various properties of jets using charged particle tracks, that are then combined using machine learning techniques. In this note a new machine learning algorithm based on graph neural networks, GN1, is introduced. GN1 uses information from a variable number of charged particle tracks within a jet, to predict the jet flavour without the need for intermediate low-level algorithms. Alongside the jet flavour prediction, the model predicts which physics processes produced the different tracks in the jet, and groups tracks in the jet into vertices. These auxiliary training objectives provide useful addition information on the contents of the jet and improve performance. GN1 compares favourably with the current ATLAS flavour tagging algorithms. For a $b$-jet tagging efficiency of $70\%$ the light ($c$)-jet rejection is improved by a factor of ~1.8 (~2.1) for jets coming from $t\bar{t}$ decays with transverse momentum $20 < p_{T} < 250$ GeV. For jets coming from $Z'$ decays with transverse momentum $250 < p_{T} < 5000$ GeV the light ($c$)-jet rejection improves by a factor ~6 (~2.8) for a comparative $30\%$ $b$-jet efficiency.