Real-Time Trigger and online Data Reduction based on Machine Learning Methods for Particle Detector Technology

Sumitted to PubDB: 2021-08-09

Category: Phd Thesis, Visibility: Public

Tags: -

Authors Steffen Baehr, Juergen Becker
Date Jan. 1, 2020
Belle II Number BELLE2-PTHESIS-2021-015
Abstract Modern particle accelerator experiments are producing immense amounts of data online during their operation. Storing the entire amount of generated data is quickly exceeding reasonable budgets for the data readout infrastructure. This problem is traditionally addressed by using a combination of trigger and data reduction mechanisms that are located close to the respective detectors to facilitate a reduction of the data rates as early in the process as possible. Meanwhile, traditional approaches to these systems are struggling with achieving an efficient reduction for modern experiments such as Belle II. The reason for this lies in the complex observed distributions of background, or unwanted, events. This situation is enhanced by the unknown characteristics of both the accelerator and detector before reaching high luminosity operation. A robust and flexible algorithmic alternative is thus required to address this problem. This can be provided by using an approach based on machine learning. Since such trigger and data reduction systems are operated under tight constraints such as small latency budgets, a high number of required data transmission connections and general real-time processing, FPGAs are used as a technological basis for the implementation. Within this thesis, several approaches based on machine learning methods were developed for FPGAs to fulfil the challenges present at the Belle II experiment. These systems are presented and discussed throughout this thesis.
Conference Karlsruhe

Files