Göttingen 2025 – wissenschaftliches Programm

Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 97: Data, AI, Computing, Electronics IX (AI-based Object Reconstruction)

T 97.3: Vortrag

Freitag, 4. April 2025, 09:30–09:45, VG 2.102

Graph Neural Networks for Track Reconstruction at the ATLAS Event Filter — •Giulia Fazzino, Sebastian Dittmeier, and André Schöning — Physikalisches Institut, Universität Heidelberg, Germany

In its High-Luminosity phase, the LHC will collide particles at unprecedented luminosity scales, drastically increasing the number of interactions per bunch crossing and thus introducing the need for upgrades in the ATLAS Trigger System. In parallel, a new tracking detector, the Inner Tracker (ITk), will be installed. Its data will be used by the Event Filter in the last step of the trigger chain, for track reconstruction and, finally, event selection.

To minimize the computing resources needed by the Event Filter, the usage of hardware accelerators such as GPUs or FPGAs is studied, and significant effort is put into the development of a tracking algorithm based on Graph Neural Networks (GNNs). Such a method would first build a graph by connecting the hits in the ITk, and subsequently generate track candidates from it thanks to a GNN and a segmentation algorithm. The construction of the graph can be conducted in several ways, one of which is to use Metric Learning, a machine learning procedure connecting hits depending on their distances in a feature space.

This talk will provide an outline of GNN-based tracking for the ATLAS Event Filter, with a focus on Metric Learning, and present results on the realization and optimization of such a graph construction method for FPGA deployment.

Keywords: Track Reconstruction; Graph Neural Networks; Event Filter