Sara McCarney ; Christian Kiesling

2019
Max-Planck-Institute for Physics Munich

Abstract: For the Belle II experiment at the SuperKEKB asymmetric electron-positron (e+/e-)collider (KEK, Japan) the concept of a first level (L1) track trigger, realized by neural networks, is presented. Using the input from a traditional Hough-based 2D track finder, the stereo wire layers of the Belle II Central Drift Chamber are used to reconstruct by neural methods the origin of the tracks along the beam (z) direction. A z-trigger for Belle II is required to suppress the dominating background of tracks from outside of the collision point. This so-called Neurotrigger is based on a Multi-Layer Perceptron (MLP)Architecture and is implemented in FPGA hardware to trigger on events in real-time, satisfying a fixed latency budget of 300 ns. The Neural Networks are trained offline in a supervised learning process using Monte Carlo (MC) particles as targets. The full L1 track trigger can be simulated in software to obtain resolutions, comparing the `true' MC values to the predicted values of the network. By means of these software simulations, one can find optimal parameters for the preprocessing and training of the Neurotrigger. This thesis presents the results of such software simulations. Resolutions of about 2 cm in the high particle transverse momentum (pt) region, and about 5 cm in the low pt region are determined, sufficient for efficient background rejection. The importance of the selected drift time input algorithm on the optimal spatial resolution of the z-trigger and trainings for a preliminary z-cut of 40 cm are discussed.

Note: Presented on 08 08 2019
Note: MSc