Physics Performance of the ATLAS GNN4ITk Track Reconstruction Chain

¹ L2IT, Laboratoire des 2 Infinis—Toulouse, Toulouse, France
² Scientific Data Division, Lawrence Berkeley National Laboratory, Berkeley, CA 94720, USA
³ Physics Department, University of Wisconsin-Madison, Madison, WI 53706, USA

^* e-mail: xju@lbl.gov

Published online: 6 May 2024

Abstract

Particle tracking is vital for the ATLAS physics programs. To cope with the increased number of particles in the High Luminosity LHC, ATLAS is building a new all-silicon Inner Tracker (ITk), consisting of a Pixel and a Strip subdetector. At the same time, ATLAS is developing new track reconstruction algorithms that can operate in the HL-LHC dense environment. A track reconstruction algorithm needs to solve two problems: track finding for building track candidates and track fitting for obtaining track parameters of those track candidates. Previously, we developed GNN4ITk, a track-finding algorithm based on a Graph Neural Network (GNN), and achieved good track-finding performance under realistic HL-LHC conditions. Our GNN pipeline relied only on the 3D spacepoint positions. This work introduces heterogeneous GNN models to fully exploit the subdetector-dependent features of ITk data, improving the performance of our GNN4ITk pipeline. In addition, we interfaced our pipeline to the standard ATLAS track-fitting algorithm and data model. With that, the GNN4ITk pipeline produces full-fledged track candidates that can be used for any downstream analyses and compared with the other track reconstruction algorithms.