Luca Schinnerl ; Thomas Kuhr ; Nikolai Hartmann

2022
LMU Munich

Abstract: The Belle II experiment is expected to accumulate a data sample of 50/ab in its lifetime. For rare processes, strong background suppression is needed to precisely measure these types of events. Because of this, an extremely large number of simulated background events is necessary for an effective analysis. However, a significant portion of the simulated data is discarded trivially in the first stage of analysis, demanding a better method of simulation to keep up with the amount of data. For this purpose a neural network is implemented to select the relevant data after the Monte Carlo event generation and then only run the costly detector simulation and reconstruction for selected events. Existing methods have shown good success with graph neural networks. However, the total speedup of simulations is limited when considering generic selections. Here a maximum speedup of the simulation with the smart background filter over the brute force method of 2.1 was reached. In this work I iteratively introduce analysis specific filters to the training of the neural networks, which can greatly increase efficiencies. For an analysis selection in the search for the rare process B -> K* nu nu this methodology has been successful in significantly improving simulation speed.

Note: Presented on 08 08 2022
Note: MSc