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

T: Fachverband Teilchenphysik

T 92: Electronics, Trigger, DAQ IV

T 92.4: Vortrag

Freitag, 20. März 2026, 09:45–10:00, KH 00.023

Low-latency GNN-based hit filtering for the Belle II Level-1 track trigger — •Greta Heine¹, Fabio Mayer², Marc Neu², Giacomo De Pietro¹, Jürgen Becker², and Torben Ferber¹ — ¹Institute of Experimental Particle Physics, Karlsruhe Institute of Technology, Karlsruhe, Germany — ²Institute for Information Processing Technology, Karlsruhe Institute of Technology, Karlsruhe, Germany

The Belle II experiment encounters increasing beam-induced background with rising instantaneous luminosity, which places tighter requirements on online and offline tracking algorithms. The Level-1 trigger must maintain a high efficiency for physics signals while operating under strict latency and bandwidth constraints. To keep trigger rates within data acquisition limits, effective background suppression, in particular in the Central Drift Chamber (CDC), is essential.

This talk presents a Graph Neural Network (GNN)-based hit filter for the Belle II Level-1 CDC hardware trigger, with a focus on the hardware-aware design workflow for FPGA deployment, including model compression via quantization-aware training and network pruning. Performance results are presented in terms of the trade-off between filtering performance and resource utilization: hit-level background rejection, impact on subsequent track reconstruction and trigger rates, as well as resource usage and sub-microsecond latency for an AMD Ultrascale FPGA implementation that meets timing after place-and-route.

Keywords: Level-1 trigger; Graph Neural Network; FPGA; data reduction; tracking