Geometric GNNs for Charged Particle Tracking at GlueX

TL;DR

This work tackles the challenge of track finding for charged particles in GlueX by using a geometry-aware Graph Neural Network (GNN) pipeline. The approach combines a graph-builder, which constructs sparse, geometry-constrained graphs (with optional skip edges to handle missing hits), and an edge-classifier GNN that predicts true track edges, enabling faster, batched GPU inference. On simulated GlueX data, the pipeline delivers a 7.5% efficiency gain at fixed purity compared to the traditional method and achieves roughly a 71% speedup in inference when processing events in batches on an A100 GPU; a smaller, FPGA-friendly model offers additional speed gains with a modest drop in performance. These results, along with FPGA timing/resource assessments, indicate strong potential for deployment in real experimental settings, with future work focusing on uncertainty quantification, track fitting, and application to higher-multiplicity environments.

Abstract

Nuclear physics experiments are aimed at uncovering the fundamental building blocks of matter. The experiments involve high-energy collisions that produce complex events with many particle trajectories. Tracking charged particles resulting from collisions in the presence of a strong magnetic field is critical to enable the reconstruction of particle trajectories and precise determination of interactions. It is traditionally achieved through combinatorial approaches that scale worse than linearly as the number of hits grows. Since particle hit data naturally form a 3-dimensional point cloud and can be structured as graphs, Graph Neural Networks (GNNs) emerge as an intuitive and effective choice for this task. In this study, we evaluate the GNN model for track finding on the data from the GlueX experiment at Jefferson Lab. We use simulation data to train the model and test on both simulation and real GlueX measurements. We demonstrate that GNN-based track finding outperforms the currently used traditional method at GlueX in terms of segment-based efficiency at a fixed purity while providing faster inferences. We show that the GNN model can achieve significant speedup by processing multiple events in batches, which exploits the parallel computation capability of Graphical Processing Units (GPUs). Finally, we compare the GNN implementation on GPU and FPGA and describe the trade-off.

Figures (10)

• Figure 1: A 3-D figure illustrating the geometry of the GlueX detector. It shows Forward Drift Chamber (FDC) that is composed of four packages, each including six disc-shaped layers. Each of the layers has detector wires on which the ($x$, $y$, $z$) coordinates of the hits are recorded through ionization energy deposits. Particles propagate along the positive $z$-axis that is perpendicular to the layers.
• Figure 2: ML Pipeline Summary: The detector yields hit coordinates information. The graph-builder constructs graph from the raw hits by introducing edges between some of the hits based on graph building constraints. Finally, the edge-classifier GNN should discard false edges only (i.e., edges not corresponding to track segments).
• Figure 3: Graphs built with no skip edges (i.e., $skip_{max} = 0$) are more susceptible to the problem of missing hits as shown in part (a). This problem is resolved by introducing edges between hits on non-consecutive layers with a maximum separation of $1 + skip_{max}$. Part (b) shows the same event built with a $skip_{max}$ of 2 that does not suffer from the gap patterns which makes detecting the full track possible (i.e., higher efficiency).
• Figure 4: Edge-Classifier GNN takes a graph (with $X$ and $E$ representing coordinates of graph hits and adjacency list, respectively) and returns the probability of each edge being true after performing $I$ message passing iterations.
• Figure 5: Overall efficiency & purity of the ML pipeline evaluated on simulated test events. At a fixed purity level of 0.9462, the pipeline achieves efficiency of 0.9806 compared to 0.9119 scored by the traditional method (7.5% increase) at a threshold $\approx$ 0.3.