Performance of the GNN-based tracking for ATLAS ITk

The Graph Neural Network (GNN) -based track finding pipeline has seen promising track reconstruction efficiencies in the ttbar events with <mu> = 200 for the ITk detector [ATL-ITK-PROC-2022-006]. The pipeline contains three discrete stages: graph construction, edge classification, and graph segmentation. The edge classification stage is the core of the pipeline and handled by a GNN. The GNN was only using the 3D global positions of space points from the Pixel and Strip detectors, neglecting the information of the two clusters associated Strip detectors. As a result, the GNN was not able to efficiently remove fake connections in the Strip barrel region. In this note, we are presenting a range of upgrade in the pipeline. First, we include the two cluster information for the Strip space points and pad the Pixel space points with their own features to reach the same length. The two types of space points are treated equally without a loss of information. The new GNN architecture uses different Neural Networks during the message-passing steps. An updated GNN per-edge efficiency and purity are presented, so are the track candidates. We compare the track reconstruction efficiency with the previous GNN tracking. Second, we designed a heterogeneous GNN model that naturally treats the Pixel and Strip space points differently. In a heterogeneous GNN, no padding is needed. We observed similar performance between the heterogeneous GNN and the default GNN. In addition, we integrated the GNN-based tracking into the offline ATLAS Athena framework, and for the first time, fit these tracks with the standard global $\chi^2$ fitter in ATLAS. The track parameter resolutions are compared between the GNN tracking and default ITk track reconstruction. We find they are compatible. The upgraded pipeline and the integration into the Athena framework make the GNN-based tracks readily usable for track-related downstream particle reconstructions and data analyses.