Mariangela Varela ; Prof. Dr. Christian Kiesling

2023
Max-Planck-Intitute for Physics Munich, Germany

Abstract: Charged pions coming from D∗ decays are very useful to tag the flavor of neutral B mesons. At the Belle II experiment, these pions have low transversal momentum and therefore, a sizeable fraction does not reach all layers of the Silicon Strip Detector (SVD). As a consequence, the tracks of these ”slow” pions are not reconstructed. This will become problematic when the Region Of Interest (ROI) algorithm is implemented online, which extrapolates the reconstructed tracks to the Pixel Vertex Detector (PXD) to select only regions of interests that are then sent to storage, deleting the rest. This thesis provides a method to recover slow pions, otherwise lost due to ROI, using only the clusters from the PXD. This PXD stand-alone cluster rescue method uses artificial neural networks (NN) to discriminate slow pions from the dominating QED electron background by analysing PXD cluster variables. The neural networks were trained using two approaches, one general and one specific: the former refers to training the NN using data from all PXD layers and all pixel multiplicities. The latter, on the other hand, refers to NNs trained with data separated according to PXD layer number and pixel multiplicity. Monte Carlo (MC) generated data was used to train and to test all NNs. Significant discrimination between slow pions and electrons has been achieved for both approaches. With the general approach, an efficiency of ∼89% was obtained. In addition, efficiencies between 84% and 99% were achieved for the specific approach, with the highest efficiencies corresponding to the cases with large pixel multiplicities (4 or more). Moreover, the algorithm was also tested with both approaches on Early Phase 3 (EP3) MC events and proved to be resilient to a more generalised background, with similar efficiencies achieved as in the case of QED electrons as background. As a last step, real slow pion data was also used for testing, giving results with a decreased performance (∼59%), mainly caused by differences between MC and real data. By considering only clusters with a charge higher than 30 ADU, MC was brought closer to real data and an efficiency of ∼73% was achieved.

Note: Presented on 07 09 2023
Note: MSc