A Real-Time Graph Neural Network Trigger Algorithm for the Belle II Electromagnetic Calorimeter

Sumitted to PubDB: 2025-09-22

Category: Phd Thesis

Tags: TRG

Principal Authors	Isabel Haide
Date	2025-07-11
Belle II Number	BELLE2-PTHESIS-2025-021
Abstract	The Belle II experiment at the SuperKEKB collider faces the challenge of efficient online event selection under high luminosity and beam background conditions. In this thesis, I present the development and implementation of a Graph Neural Network (GNN)-based algorithm for the Electromagnetic Calorimeter (ECL) Level-1 Trigger. The network, called GNN-ETM, uses dynamic graph building together with the Object Condensation approach to reconstruct an unknown number of clusters in the ECL in real time. A hardware-software co-design strategy for the network design is used to meet the strict latency and throughput requirements of an FPGA-based trigger system. The GNN-ETM shows improved efficiency, position resolution, and low-energy cluster performance compared to the current L1 ECL trigger, and introduces a signal/background classification that reduces background rates. Additionally, it has been included in the Belle II L1 trigger during collision data taking in December 2024. This work demonstrates the design and integration of a GNN-based real-time trigger algorithm with dynamic graph building in a realistic collider environment, performing within the throughput constraints.