Heidelberg 2022 – scientific programme

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

T: Fachverband Teilchenphysik

T 79: Data Analysis, Information Technology and Artificial Intelligence 4

T 79.3: Talk

Wednesday, March 23, 2022, 16:45–17:00, T-H38

Fast Particle Reconstruction in the Belle II Experiment with Graph Neural Networks — •Isabel Haide, Pablo Goldenzweig, and Torben Ferber for the Belle II collaboration — Karlsruhe Institute of Technology

The correct clustering and reconstruction of particles in electromagnetic calorimeters are vital to many analyses to ensure a correct reconstruction of the actual event. This clustering poses difficulties such as an unknown number of particles in the calorimeter and the existence of background events and promotes the use of machine learning (ML) algorithms. Due to the irregular geometry of such detectors, graph neural networks (GNNs) are most suitable to provide an improvement over standard algorithms. GNNs do not depend on regular geometries to learn detector-hit representations and have already been successfully applied to simulated data of a simplified calorimeter model. Extending this application to the geometry of real detectors, such as the Belle II electromagnetic calorimeter (ECL), while reconstructing an unknown number of clusters with possible overlap and additional background events, is the goal of this study. In this talk, the concept of using object condensation with GNNs to reconstruct particles in the ECL and the current status of this development is shown. The evaluation method, which is the separation of the signature of the hypothetical dark photon process e⁺ e⁻ → A′ γ, A′ → e⁺ e⁻ to the signature of radiative Bhabha scattering e⁺ e⁻ → e⁺ e⁻ γ, is also explained.