Transformer Neural Networks for Identifying Boosted Higgs Bosons decaying into $b\bar{b}$ and $c\bar{c}$ in ATLAS

Identifying boosted Higgs bosons decaying hadronically into a pair of $b$-quarks or $c$-quarks is an important capability which opens many opportunities to enrich the ATLAS physics programme at the Large Hadron Collider. In this work, a new algorithm using a transformer neural network architecture is introduced, GN2X, for identifying large-radius jets originating from boosted Higgs bosons decaying to $b\bar{b}$ and to $c\bar{c}$ pairs. GN2X directly uses information from charged particle tracks associated with the large-radius jet. For a working point exhibiting 50% $H(b\bar{b})$ efficiency, GN2X achieves a background rejection factor of 40 for jets from top-quark decays and 300 for multijet events. GN2X offers significant improvements in the rejection of background jets over the previous approach, which uses flavour tagging discriminants of individual track-based subjets in a feed-forward neural network architecture. Additional GN2X variants combining tracks, subjets and large-radius jet calorimeter constituents are also explored and found to achieve further gains in background rejection.