Tagging more quark jet flavours at FCC-ee at 91 GeV with a transformer-based neural network

Abstract

Jet flavour tagging is crucial in experimental high-energy physics. A tagging algorithm, DeepJetTransformer, is presented, which exploits a transformer-based neural network that is substantially faster to train than state-of-the-art graph neural networks. The DeepJetTransformer algorithm uses information from particle flow-style objects and secondary vertex reconstruction for $b$- and $c$-jet identification, supplemented by additional information that is not always included in tagging algorithms at the LHC, such as reconstructed $K_S^0$ and ${\Lambda }^0$ and $K^{\pm }/{\pi }^{\pm }$ discrimination. The model is trained as a multiclassifier to identify all quark flavours separately and performs excellently in identifying $b$- and $c$-jets. An $s$-tagging efficiency of $40\%$ can be achieved with a $10\%$ $ud$-jet background efficiency. The performance improvement achieved by including $K_S^0$ and ${\Lambda }^0$ reconstruction and $K^{\pm }/{\pi }^{\pm }$ discrimination is presented. The algorithm is applied on exclusive $Z\to q\bar{q}$ samples to examine the physics potential and is shown to isolate $Z\to s\bar{s}$ events. Assuming all non-$Z\to q\bar{q}$ backgrounds can be efficiently rejected, a $5\sigma$ discovery significance for $Z\to s\bar{s}$ can be achieved with an integrated luminosity of $60{\text{nb}}^{-1}$ of $e^{+}{e}^{-}$ collisions at $\sqrt{s}=91.2\mathrm{GeV}$, corresponding to less than a second of the FCC-ee run plan at the $Z$ boson resonance.