Lorentz-Equivariant Geometric Algebra Transformers for Top-Quark Event Reconstruction

Results are presented for Lorentz-equivariant geometric algebra transformers (LGaTr) in three event reconstruction tasks involving top quarks. Geometric algebra provides a common Lorentz-covariant representation of four-momenta and typed physical objects. Performance of LGaTr is evaluated on three different physics tasks - $t\bar{t}\mathrm{/}tWb$ separation, FCNC detection in single top-quark production and universal event reconstruction. LGaTr is outperforming the baseline networks in $tW\mathrm{/}t\bar{t}$ separation and full event reconstruction and is showing comparable or better performance to standard transformer-based networks. The implementation is provided in a companion repository https://github.com/emil2001/top-quark-universal-nets.