Erlangen 2026 – scientific programme
Parts | Days | Selection | Search | Updates | Downloads | Help
T: Fachverband Teilchenphysik
T 28: Data, AI, Computing, Electronics III
T 28.2: Talk
Tuesday, March 17, 2026, 16:30–16:45, KH 00.024
Graph Neural Networks for multi-hypothesis clustering in the Belle II Electromagnetic Calorimeter to improve hadron clustering — •Jonas Eppelt and Torben Ferber — Karlsruhe Institute for Technology
The Belle II experiment at the SuperKEKB collider in Tsukuba, Japan, studies the products of e+ e− collisions to probe the Standard Model and search for new physics. Many of these processes involve π0, which almost always decay into two γ, and reconstructing these correctly is Important for many analyses. The currently used clustering algorithm is optimized towards the reconstruction of photon clusters, which form regular, mostly round clusters. However, the e+ e− collisions also produce hadronic particles, which also interact in the calorimeter. As their energy depositions are more irregularly shaped and can produce split-off particles, which create disconnected, additional clusters, they pose a significant challenge for any clustering optimized for electromagnetic clusters. Improving upon their reconstruction would not only improve their identification, but also help constrain the energy in the calorimeter to the collision energy. I will show an implementation of a graph neural network, optimized for both photon and hadron reconstruction. Not only does it improve the photon energy resolution, but it also improves the hadron position reconstruction. Further, I will demonstrate the improved constraints on the energy in the calorimeter on the collision energy.
Keywords: Graph Neural Networks; Reconstruction; Clustering