A Fast AI-Based Track Reconstruction on FPGA for the PANDA Experiment

Efficient reconstruction of charged particle trajectories is a crucial yet very difficult step in the analysis pipeline of high energy physics (HEP) experiments. Recent work has shown that graph neural networks (GNNs) are well suited for the pattern recognition task of track finding, where tracking detector hits can be naturally represented as nodes and particle track segments as edges. The interaction network (IN) GNN architecture provides computationally efficient edge classification of the high-dimensional and sparse tracker data, which is especially crucial for implementation in constrained computing environments such as field programmable gate arrays (FPGAs). This work describes the overall workflow for implementing and systematically analyzing an IN-based classification of track segments on FPGAs for the anti-Proton Annihilation at DArmstadt (PANDA) forward tracking system (FTS). This workflow includes data preprocessing, graph building, GNN-based edge classification and a series of FPGA implementation design studies concerning latency, resource utilization, and classification quality using the high level synthesis for machine learning (hls4ml) compiler. The presented final implementation of the GNN-based track segment classifier on a Xilinx Zynq provides an overall inference latency of about 0.99 μs using about 34 % of available digital signal processors (DSPs) and 85 % of available lookup tables (LUTs). This work enables the acceleration of charged particle tracking on heterogeneous computational resources toward real-time track reconstruction for the PANDA experiment. The discussed methods and studies could be easily adapted and used in other HEP experiments for accelerated charged particle tracking.