Graph Neural Networks for Charged Particle Tracking on FPGAs

Area of Science:

• High Energy Physics
• Machine Learning
• Computer Engineering

Background:

• Accurate charged particle tracking is essential for analyzing high-energy collisions at the Large Hadron Collider (LHC).
• Future high-luminosity LHC (HL-LHC) conditions present significant challenges due to high interaction densities.
• Graph neural networks (GNNs) show promise for particle tracking but face computational cost limitations in real-time trigger applications.

Purpose of the Study:

• To develop an efficient method for implementing GNNs for charged particle tracking on Field-Programmable Gate Arrays (FPGAs).
• To enable the use of GNNs in hardware-based trigger systems for the HL-LHC.
• To address the computational challenges of GNNs in high-energy physics experiments.

Main Methods:

• An automated translation workflow within the hls4ml tool was developed to convert GNNs into FPGA firmware.
• GNNs for charged particle tracking were trained using the TrackML challenge dataset.
• Implementations targeted various graph sizes, task complexities, and latency/throughput requirements on FPGAs.

Main Results:

• Successful implementation of GNNs for charged particle tracking on FPGAs was demonstrated.
• The developed workflow allows for efficient conversion of GNN models for hardware deployment.
• The performance of FPGA implementations was evaluated across different design parameters.

Conclusions:

• The automated translation workflow facilitates the integration of GNNs into FPGA-based trigger systems for the HL-LHC.
• This approach can significantly enhance real-time data processing capabilities in high-energy physics.
• Enabling GNNs at the trigger level opens new possibilities for physics discovery at future colliders.