Title: Enhancing the Graph Neural Networks Inference for Track Reconstruction in High-Energy Physics

Presenter: Alina Lazar, Professor, Youngstown State University

Abstract:

Graph Neural Network (GNN) models proved to perform well on the particle track reconstruction problem, where traditional algorithms become computationally complex as the number of particles increases, limiting the overall performance. Track finding is a crucial step in analyzing data from particle detectors, like those used in particle accelerators, where it involves identifying and reconstructing the trajectories (tracks) of charged particles produced in high-energy collisions. GNNs can capture complex relationships in event data represented as graphs. However, training and especially inference on large graphs is challenging due to computation and GPU memory requirements. Fast inference speeds in GNN-based track reconstruction are crucial for managing large data volumes, ensuring timely decision-making, optimizing resource usage, enabling more complex physics analyses, and maintaining scalability in high-energy physics experiments. The talk is dedicated to discussing the implementation of hardware-accelerated computing, particularly the use of GPUs and CPUs, in optimizing the GNN-based pipeline for inference. We present case studies demonstrating the impact of these optimizations on processing speed, accuracy, and computing resource utilization.