Timo Forsthofer ; Prof. Christian Kiesling

2024
Max-Planck-Intitute for Physics Garching

Abstract: In order to provide accurate measurements of Standard Model parameters and search for new physics at the precision frontier [1], the Belle II experiments aims to reach an instantaneous luminosity of 6 × 1035 cm−2s−1 [2]. An important tool for coping with the resulting high background levels is the Neural Track Trigger (NTT). This first level trigger makes use of a neural network to predict the origin of a particle along the beam (z) direction, rejecting tracks with a large displacement [3]. In previous running periods, it has been very successful in triggering particularly low multiplicity events with a high efficiency while maintaining a low trigger rate. However, during high luminosity runs before the Long Shutdown (LS1) in June 2022, the trigger rate drastically increased, making an upgrade necessary. This thesis studies how new computing hardware and new software developments can be utilized for optimizing the performance of the NTT. As a first step, different deep neural network architectures within the hardware restrictions are tested. Next, the number of input nodes is increased to include more information, thereby enhancing the z resolution. Using an ADC-cut and the 3DFinder described in [4] and [5], background and fake tracks are suppressed and the quality of the input parameters is improved. Finally, an output node for classification is added, replacing the z-cut. All of this significantly improves both the single-track efficiency and the background rejection rate, making the trigger stable even at the high background rates expected in the future.

Note: Presented on 30 06 2024
Note: MSc