Combined track finding with GNN & CKF

TL;DR

Under HL-LHC conditions with pile-up up to $\\langle \\mu \\rangle \\approx 200$, track reconstruction challenges escalate for traditional algorithms. The authors propose a hybrid pipeline that seeds tracks with a Graph Neural Network (GNN) in the inner, high-resolution pixel region and uses a Combinatorial Kalman Filter (CKF) for outer regions, leveraging CKF’s natural handling of 1D measurements and reducing graph complexity. Evaluated within the ACTS framework on OpenDataDetector using $t\\bar{t}$ events with PU 200, the approach achieves target-edge purity $>0.99$ after the second GNN and shows improved performance when short-strip barrel hits are included, with per-event times around $\\approx 2$ s on a single Nvidia A100. This work demonstrates a viable, hardware-friendly pathway toward scalable track reconstruction for HL-LHC-like conditions and informs design choices on seed generation and fit-stage integration.

Abstract

The application of Graph Neural Networks (GNN) in track reconstruction is a promising approach to cope with the challenges arising at the High-Luminosity upgrade of the Large Hadron Collider (HL-LHC). GNNs show good track-finding performance in high-multiplicity scenarios and are naturally parallelizable on heterogeneous compute architectures. Typical high-energy-physics detectors have high resolution in the innermost layers to support vertex reconstruction but lower resolution in the outer parts. GNNs mainly rely on 3D space-point information, which can cause reduced track-finding performance in the outer regions. In this contribution, we present a novel combination of GNN-based track finding with the classical Combinatorial Kalman Filter (CKF) algorithm to circumvent this issue: The GNN resolves the track candidates in the inner pixel region, where 3D space points can represent measurements very well. These candidates are then picked up by the CKF in the outer regions, where the CKF performs well even for 1D measurements. Using the ACTS infrastructure, we present a proof of concept based on truth tracking in the pixels as well as a dedicated GNN pipeline trained on $t\bar{t}$ events with pile-up 200 in the OpenDataDetector.

Figures (6)

• Figure 1: Target-edge efficiency of the different stages as a function of $\eta$, $p_T$ and the number of pixel hits of the corresponding particle for the pipeline trained with the pixel hits only
• Figure 2: Same as \ref{['fig:2gnn_edge_eff']} but trained with the pixel hits and the hits in the short-strip barrel.
• Figure 3: Efficiency, duplication rate, and fake rate vs particle $\eta$ and $p_T$, evaluated with 50 events.
• Figure 4: Detailed matching efficiency for two versions of the GNN+CKF chain, evaluated with 50 events.
• Figure 5: Metrics for the GNN+CKF chain with and without the combinatorial aspect in the CKF stage.