Stella Katharina Wermuth ; Dr. Torben Ferber ; Prof. Dr. Oliver Gerberding

2021
DESY Hamburg

Abstract: The Belle II experiment, located at the SuperKEKB e+e- collider in Japan, uses pulse shape analysis techniques to distinguish electromagnetically and hadronically interacting particles within the CsI(Tl) electromagnetic calorimeter. The pulse shapes from the particle-dependent scintillation response are nominally analysed with a multi-template offline fit to measure the fraction of scintillation emission produced by hadrons. This fitting method allows for the determination of the total deposited energy, the total scintillation emission by hadrons, and the time of energy deposit. This thesis reports on a new approach to extract the total deposited energy, and the hadronic component of the scintillation emission from the pulse shapes using machine learning techniques. For this, a neural network is trained on pulse shapes produced in crystals from calorimeter clusters from simulated photons and pions, and is employed as a multivariate regression tool. Inferred on photons, the neural network outperforms the current fitting method in terms of crystal energy resolution and hadron intensity resolution. For pions the neural network shows a similar resolution compared with the current fitting method. Furthermore the neural network approach improves the discrimination of electromagnetic and hadronic interactions and is robust towards fluctuations in photon pile-up from beam backgrounds. Overall the neural network approach is promising, however additional fine tuning of the composition of the training sample could further improve its performance and robustness.

Note: Presented on 22 06 2021
Note: MSc