Multi-Modal Track Reconstruction using Graph Neural Networks at Belle II

TL;DR

The paper addresses degraded track reconstruction at Belle II due to backgrounds and detector ageing by introducing BAT Finder, a unified graph neural network that jointly processes SVD and CDC hits in a single inference. Leveraging GravNet blocks and the Object Condensation loss, it reconstructs full tracks without detector-specific staging, achieving a track efficiency of 74.7% and purity of 97.63% for displaced muons—significant improvements over the Baseline and prior CAT approaches. The method relies on full detector simulations with beam backgrounds and demonstrates robust performance across diverse topologies, suggesting strong potential for sustaining high-precision tracking in high-luminosity Belle II operations. This work advances end-to-end, multi-modal track reconstruction by integrating heterogeneous detector inputs and reducing misreconstruction, with practical impact on vertexing, momentum determination, and background suppression.

Abstract

High backgrounds and detector ageing impact the track finding in the Belle II central drift chamber, reducing both track purity and track efficiency in events. This necessitates the development of new track finding algorithms to mitigate detector performance degradation. Building on our previous success with an end-to-end multi-track reconstruction algorithm for the Belle II experiment at the SuperKEKB collider (arXiv:2411.13596), we have extended the algorithm to incorporate inputs from both the drift chamber and the silicon vertex tracking detector, creating a multi-modal network. We employ graph neural networks to handle the irregular detector structure and object condensation to address the unknown, varying number of particles in each event. This approach simultaneously identifies all tracks in an event and determines their respective parameters. We demonstrate the algorithm's effectiveness using a realistic full detector simulation, which incorporates beam-induced backgrounds and noise modelled from actual collision data. The simultaneous reconstruction of the information from the two detectors yields a track efficiency improvement from 48.0 % to 74.7 % for uniformly displaced particles up to 100 cm, while increasing the track purity by 5.5 percentage points. We provide a detailed comparison of its track-finding performance against the current Belle II baseline across various event topologies.

Figures (3)

• Figure 1: The Belle II track finding chain for high level trigger tracks and tracks used in offline reconstruction. The blue box shows the CDC track finding algorithm, while the purple box indicates the SVD track finding algorithm. Final reconstruced objects are highlighted in orange. The inputs of the three tracking detectors CDC, SVD, and PXD are given by one example event display, where the hits from signal particles are shown in colored circular markers and the beam-background is shown with grey triangular markers. See text for details.
• Figure 2: BAT Finder model architecture.
• Figure 3: Track efficiency as a function of transverse displacement $v_{\rho}^{\mathrm{MC}}$ for the Baseline Finder in blue, CAT Finder in orange and BAT Finder in red. The grey vertical line marks the transition between the SVD and the CDC, corresponding to the outer radius of the SVD. For displacements beyond this radius, tracks originate inside the CDC and no longer traverse the SVD.