Characterization and upgrade of a quantum graph neural network for charged particle tracking
Matteo Argenton, Laura Cappelli, Concezio Bozzi · 2026
Read Characterization and upgrade of a quantum graph neural network for charged particle tracking by Matteo Argenton, Laura Cappelli, Concezio Bozzi (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-08667.txt.
Introduction
Matteo Argenton and colleagues characterize and upgrade a quantum graph network that rebuilds charged particle tracks at the LHC. Brighter, denser collisions will tangle track finding, so the team tests a hybrid design mixing plain feedforward nets with tuned quantum circuits. They map how the plain and quantum parts interact and report smoother training toward the final fit.
Catalog record
| CREATOR | Matteo Argenton, Laura Cappelli, Concezio Bozzi |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 44622 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.08667
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Characterization and upgrade of a quantum graph neural network for charged particle tracking Matteo Argenton, Laura Cappelli, Concezio Bozzi Abstract In the forthcoming years the LHC experiments are going to be upgraded to benefit from the substantial increase of the LHC instantaneous luminosity, which will lead to larger, denser events, and, consequently, greater complexity in reconstructing charged particle tracks, motivating frontier research in new technologies. Quantum machine learning models are being investigated as potential new approaches to high energy physics (HEP) tasks. We characterize and upgrade a quantum graph neural network (QGNN) architecture for charged particle track reconstruction on a simulated high luminosity dataset. The model operates on a set of event graphs, each built from the hits generated in tracking detector layers by particles produced in proton collisions, performing a classification of the possible hit connections between adjacent layers. In this approach the QGNN is designed as a hybrid architecture, interleaving classical feedforward networks with parametrized quantum circuits. We characterize the interplay between the classical and quantum components. We report on the principal upgrades to the original design, and present new evidence of improved training behavior, specifically in terms of convergence toward the final trained configuration. I Introduction Within high energy physics (HEP), the reconstruction of the trajectories of charged particles, starting from the clustered energy deposits left during their propagation through the tracking layers of a detector, is a computationally intensive, but central task 5 . At the luminosity regimes that the Large Hadron Collider (LHC) is entering in 2030, which marks the beginning of its high luminosity phase (HL-LHC) 4 , particle tracking in the two major general purpose experiments, CMS and ATLAS, will need to account for an increment of a factor 3-5 in the average number of primary proton-proton inelastic interactions (referred to as pileup ), averaging 140–200 in future runs 15 . Similar upgrades, although targeting different levels of luminosity, are also planned for the LHCb 34 and ALICE 20 detectors at a later stage. While the tracking pipelines currently in use up to Run 3 17 ; 43 ; 26 ; 1 are being upgraded to handle the combinatorial blow-up, which is a consequence of the increased proton flux, through e.g. pileup mitigation techniques 8 ; 46 , these tracking stacks are often rooted in the combinatorial Kalman Filter (CKF) algorithm 9 ; 37 , and are therefore inherently challenged because of their local approach to track reconstruction. In recent years, a plethora of novel approaches has emerged, mostly based on machine learning models, and targeting different stages of the reconstruction pipeline. In particular, graph neural networks (GNN) are one of the most promising technologies, both specifically for particle tracking 31 ; 16 ; 19 , and for a broader class of particle physics applications 23 ; 42 ; 6 . Alongside these ongoing developments in machine learning for high energy physics, other novel computing paradigms are being researched. In particular, quantum technologies are explored as alternative, and often synergic, computing paradigms for solving classically inefficient tasks. Within this scope, quantum machine learning (QML) is a hybrid field targeting to understand how learning from data can be performed through quantum algorithms, and it has seen a considerable development in the last decade, often in parallel with, but not constrained to, modern machine learning models. The research presented in this work focuses on a QML application to the specific task of track segment classification in particle tracking. We discuss the development, characterization, and upgrade of a quantum graph neural network (QGNN) model first proposed in 47 , and here extended to provide improved training and generalization performance. QGNNs belong to the broad class of variational quantum machine learning algorithms that can be tested in the current NISQ (Noisy Intermediate Scale Quantum) era, where the number of qubits and circuit depth are limited both for quantum hardware, and classical simulators. In particular, QGNNs are an advanced application of parametrized quantum circuits (PQC), to classical graph neural networks. This class of architectures exploits the flexibility and representation capacity of classical multi-layer perceptrons (MLP), to feed and extract data from the circuits. This study is aimed at improving the understanding of hybrid variational quantum algorithms, and their possible applications in experimental particle physics tasks. II Dataset and event graph construction The high pileup data used in this research comes from the simulated TrackML dataset 3 , which is a standard reference for machine learning research on particle tracking. The dataset is based on a modular, synthetic tracking detector whose design is based on the high-luminosity upgrades of both the CMS and ATLAS inner trackers; additional details on the detector are reported in Appendix A . Before being used as input to the QGNN, the TrackML events undergo a preprocessing step, consisting of a set of constraints on the hits’ physical coordinates – necessary to reduce the number of hits and, consequently, of candidate track segments – and a graph-building stage, where an input graph is built from the selected hits of each collision event. II.1 Preprocessing constraints In the tests presented in this work only the hits in the central (barrel) regions of Fig. 1 are considered; for a total of ten tracking planes. This restriction can be lifted in future studies, but represents a good compromise, especially in the hybrid quantum framework, between the time needed to train the models, and the fidelity to a real-world tracking scenario. Similarly, a cut is enforced on the particles’ transverse momentum, to only retain p_{T}\geq 1 GeV tracks. This cut is also motivated by the training time of the hybrid quantum model, both when the quantum resources are simulated on classical hardware, and when tests are performed on quantum hardware. In this research the geometrical constraints on the hits are aligned with the ones of the original QML study 47 , and of the classical HEP.TrkX project 24 from which it originates. The coordinates of each hit are expressed in a cylindrical reference system (r,\phi,z) , where r is the radial distance from the beam axis z , and \phi is the azimuthal angle. For each hit pair (h_{1},h_{2}) the following quantities are computed and constrained: • \phi_{\text{slope}}=\Delta\phi/\Delta r • z_{0}=z_{1}-r_{1}\Delta z/\Delta r with cuts: • |\phi_{\text{slope}}|<6\times 10^{-4} • \left|z_{0}\right|<100\mathrm{~mm} . Figure 1: TrackML detector layout from 32 . The three different sub-detectors are shown on the left and with the associated colors in the z-r plane depiction on the right. Grey lines correspond to different pseudorapidity \eta values. II.2 Graph building A graph \mathcal{G}=(\mathcal{V},\mathcal{E}) is a collection of objects \mathcal{V} (named nodes, or vertices), together with a set of relations between said objects \mathcal{E} (edges, or arcs). Given two nodes i,j\in\mathcal{V} ; the edge going from i to j is e=(i,j)\in\mathcal{E} . If the edge has no preferential direction, hence (i,j)\equiv(j,i) the edge is said to be undirected; if the opposite is true it is a directed edge. This work considers only directed edges because they have shown to consistently outperform models of the latter kind. Edges can be both weighted or unweighted. An arbitrary graph is weighted if it is equipped with a weight function w:\mathcal{E}\to\mathbb{R} , where the weights w_{e} are all non trivially one. In