Design and Evaluation of a GNN Algorithm for the ECL L1 Trigger Upgrade at Belle II

One of the central challenges of the Belle II experiment at the SuperKEKB collider is to further increase the instantaneous luminosity in order to enlarge the available dataset for precision measurements in particle physics. However, due to the low cross-section of relevant processes, only a small fraction of bunch crossings result in meaningful collisions. Most recorded events are dominated by background processes such as Bhabha scattering, which are not of primary interest. Efficient event filtering is therefore essential.

This filtering is performed through a two-stage trigger system consisting of a hardware-based Level-1 (L1) trigger and a software-based High-Level Trigger (HLT). The L1 trigger must make real-time decisions using FPGA-based algorithms to reduce the event rate to a level manageable by the HLT, while preserving physics-relevant events. With the planned luminosity increase during Long Shutdown 2 (LS2), trigger rates are expected to exceed current limits, posing significant challenges for the existing trigger infrastructure.

In this work, a novel high-granularity clustering approach for the Electromagnetic Calorimeter (ECL) L1 trigger is presented. Building on the existing Graph Neural Network-based CaloClusterNet, this study explores adaptations enabled by upcoming hardware upgrades that allow the use of single-crystal information. Two input reduction schemes and corresponding model variants are developed and evaluated, with comparisons to the existing lower granularity clustering algorithms and the offline reconstruction. Additionally, the feasibility of real-time deployment is demonstrated through a hardware implementation.