Parts | Days | Selection | Search | Updates | Downloads | Help

T: Fachverband Teilchenphysik

T 28: Data, AI, Computing, Electronics III

T 28.3: Talk

Tuesday, March 17, 2026, 16:45–17:00, KH 00.024

Point Cloud Segmentation for the Belle II GNN-Based Tracking — •Daniel Grossmann, Tristan Brandes, Giacomo De Pietro, Torben Ferber, and Lea Reuter — Institute of Experimental Particle Physics, Karlsruhe Institute of Technology, Karlsruhe, Germany

Our implementation of an end-to-end multi-track based reconstruction algorithm for the Belle II experiment at the SuperKEKB collider improves the tracking performance compared to the baseline algorithm (arXiv:2411.13596). It combines the Object Condensation algorithm with a Graph Neural Network that simultaneously identifies all tracks in an event and determines their respective parameters. However, our current algorithm is based on a segmentation step during the model post-processing, which fails to capture specific signatures for more complex track topologies.

This work improves the segmentation by employing a pointcloud-based model in the post-processing step for the track and hit assignment. We report the performance of our improved track-segmentation algorithm across various event topologies compared to the existing segmentation method and to the baseline tracking algorithm used in Belle II.

Keywords: graph neural networks; pointcloud; reconstruction; machine learning; clustering