Bereiche | Tage | Auswahl | Suche | Aktualisierungen | Downloads | Hilfe

T: Fachverband Teilchenphysik

T 54: Data, AI, Computing, Electronics V (Anomaly Detection, Event Selection)

T 54.3: Vortrag

Mittwoch, 2. April 2025, 16:45–17:00, VG 2.101

BGNet: A neural network for real-time background prediction and decomposition for Belle II — •Yannik Buch, Ariane Frey, Benjamin Schwenker, and Lukas Herzberg — Georg-August-Universität Göttingen, Göttingen, Deutschland

The Belle II detector investigates the b-sector by measuring the decays of the Υ(4S) resonance. These resonances are produced by the SuperKEKB accelerator at KEK in Tsukuba, Japan. The goal of SuperKEKB is to achieve an instantaneous luminosity of 6.5 × 10³⁵ cm⁻²s⁻¹. Currently, a luminosity of 5 × 10³⁴ cm⁻²s⁻¹ is reached, showing that considerable improvements to the beam focusing and increases of the ring currents are still necessary. At the same time, however, the Belle II detector must not be damaged or its performance compromised by extensive radiation and hit rates. The beam backgrounds at Belle II are mostly composed of storage backgrounds, luminosity-based backgrounds and injection backgrounds of both rings due to continuous top-up injections. BGNet is trained to predict the overall hit rates and their decomposition in terms of background source for various Belle II sub-detectors.

The input data for BGNet are 1 Hz time series of diagnostic variables describing the state of the SuperKEKB collider subsystems. Using real-time data from the EPICS slow control system BGNet can be used to obtain a real-time beam background decomposition, enabling diagnostic background monitoring for all beam background components simultaneously.

Keywords: Neural Networks; Machine Learning; Belle II; SuperKEKB; Beam backgrounds