Logo Logo
FAQ
Contact
Switch language to German
Full event reconstruction with transformers and hyperbolic embedding, and search for the rare decays B → πνˉv at the Belle II Experiment
Full event reconstruction with transformers and hyperbolic embedding, and search for the rare decays B → πνˉv at the Belle II Experiment
In this work, I first present HyperTagging, a machine-learning framework for full-event reconstruction of generic decays. The method employs transformer-based architectures and representation learning techniques to embed reconstructed particles and decay structures into a learned latent space. By exploiting similarities between decay topologies, the algorithm provides a flexible approach for reconstructing complex decay chains and can improve the efficiency of identifying the root particle in events with missing energy. The performance of the method is evaluated on simulated datasets and exceeds that of baseline models. Second, I perform a sensitivity study of the rare decays B → πνˉv, which proceed via the FCNC transition b → dνˉv. The analysis is based on simulated BelleII datasets corresponding to an integrated luminosity of 492 fb⁻¹ and uses exclusive tag-side reconstruction. The best upper limit is obtained in the B⁺ → π⁺ νˉv channel with hadronic tag-side reconstruction, reaching $1.04 × 10⁻⁵ at 90% C.L., which is about 26% better than the previous best result.
Not available
Yu, Boyang
2026
English
Universitätsbibliothek der Ludwig-Maximilians-Universität München
Yu, Boyang (2026): Full event reconstruction with transformers and hyperbolic embedding, and search for the rare decays B → πνˉv at the Belle II Experiment. Dissertation, LMU München: Faculty of Physics
[thumbnail of Yu_Boyang.pdf]
Preview
PDF
Yu_Boyang.pdf

14MB

Abstract

In this work, I first present HyperTagging, a machine-learning framework for full-event reconstruction of generic decays. The method employs transformer-based architectures and representation learning techniques to embed reconstructed particles and decay structures into a learned latent space. By exploiting similarities between decay topologies, the algorithm provides a flexible approach for reconstructing complex decay chains and can improve the efficiency of identifying the root particle in events with missing energy. The performance of the method is evaluated on simulated datasets and exceeds that of baseline models. Second, I perform a sensitivity study of the rare decays B → πνˉv, which proceed via the FCNC transition b → dνˉv. The analysis is based on simulated BelleII datasets corresponding to an integrated luminosity of 492 fb⁻¹ and uses exclusive tag-side reconstruction. The best upper limit is obtained in the B⁺ → π⁺ νˉv channel with hadronic tag-side reconstruction, reaching $1.04 × 10⁻⁵ at 90% C.L., which is about 26% better than the previous best result.