Training and Optimization of a Graph Neural Network for Deployment on FPGA Hardware in the Belle II Level 1 Trigger

At the Belle II experiment, collisions occur at a rate higher than 200 MHz. Due to bandwidth and storage limitations, storing this amount of data for offline analysis is not feasible. That is why the experiment relies on a two-level trigger system as a preselection before passing the events to subsequent software-based processing and analysis. The first level, known as the Level 1 Trigger (L1 Trigger), is a hardware-based system that makes decisions within a few microseconds. The buffered data from the detector can only be stored for a limited time, which imposes strict latency requirements on the L1 Trigger.

One approach to enhance the L1 Trigger is the development of the GNN-ETM. It utilizes dynamic graph building based on GravNet and can predict a varying number of clusters per event using Object Condensation. The GNN-ETM is implemented on field-programmable gate array (FPGA) hardware, which allows for low-latency inference and is suitable for deployment on the L1 Trigger. The current parasitic implementation of the GNN-ETM in the L1 Trigger has shown promising results in terms of precision and latency. Nevertheless, further improvements, especially in terms of latency, are necessary to develop a potential replacement for the clustering algorithm currently deployed on the ECL L1 Trigger.

In this thesis, I present the optimization of the GNN-ETM for reduced latency, including modifications to the model architecture and training procedure.