Abstract

A neural network based z-vertex trigger is developed for the first level trigger of the upgraded flavor physics experiment Belle II at the high luminosity B factory SuperKEKB in Tsukuba, Japan. Using the hit and drift time information from the central drift chamber, a pool of expert neural networks estimates the 3D track parameters of the single tracks found by a 2D Hough finder. The neural networks are already implemented on parallel FPGA hardware for real time data processing and running pipelined in the online first level trigger of Belle II. Due to the anticipated high luminosity of up to 8 × 10³⁵ cm⁻²s⁻¹, Belle II will have to face severe levels of background tracks with vertices displaced along the beamline. The neural z-vertex algorithm presented in this thesis allows to reject displaced background tracks such that the requirements of the standard track trigger can be strongly relaxed. Especially for physics decay channels with a low track multiplicity in the final states, like τ pair production, or initial state radiation events with reduced center of mass energies, the trigger efficiencies can be significantly increased. As an upgrade of the present 2D Hough finder in the neural network preprocessing, a model independent 3D track finder is developed that uses the additional stereo hit information of the drift chamber. Thus, the trigger efficiencies improve for tracks in the phase space of low transverse momenta and shallow polar angles. Since the cross sections of the physics signal events typically increase towards shallow polar angles, this enlarged acceptance of the track trigger provides a substantial gain in the signal efficiencies. By using an adapted pool of expert networks, the enlarged phase space provided by the 3D finder can be efficiently covered. Studies on simulated MC background, on simulated initial state radiation events, and on recorded data from early Belle II runs demonstrate the high performance of the novel trigger algorithms. With the 3D finder an increase of the track finding rate of about 50 % is confirmed for signal tracks, while displaced background tracks are actively suppressed prior to the neural network. Based on z-vertex cuts on the tracks processed by the neural networks, a two track event efficiency of more than 99 % can be achieved with a purity of around 80 %.