Investigation of the performance of a GNN-based b-jet tagging method in heavy-ion collisions

Abstract

Beauty-tagged jets (b-jets)-collimated sprays of particles originating from the fragmentation of beauty quarks produced in the initial hard scatterings-provide a unique probe of parton dynamics in the quark-gluon plasma (QGP) created in ultrarelativistic heavy-ion collisions. In particular, energy loss patterns of low-$p_T$ b-jets traversing the QGP offer valuable insight into the strong interaction in its nonperturbative regime. CMS and ATLAS Collaborations at the LHC have studied b-jet production in Pb-Pb collisions. The results were limited to a high-$p_T$ region, because a major challenge at low-$p_T$ is the overwhelming number of background particles from QGP hadronisation, which severely hinders the effectiveness of conventional b-jet tagging techniques. To enable precise measurements in such complex environments, advanced tagging methods are required. Graph Neural Networks (GNNs), capable of learning relational structures among jet constituents, represent a promising deep learning approach for b-jet identification. In this study, we adopt and adapt the GN1 model, initially developed by ATLAS, for use in Pb-Pb collision environments. We investigate the model's performance by applying it to jets embedded with Pb-Pb background particles, evaluating both tagging decisions and robustness against background contamination. This work presents a comprehensive evaluation of GNN-based b-jet tagging under heavy-ion collision conditions, aiming to advance future precision studies of QGP-induced partonic energy loss.