Corentin LEMOINE ; Jerome Baudot ; Luca Federici

2022
IPHC Strasbourg

Abstract: An efficient trigger system is essential in most particle physics experiment to cope with the conjunction of high data rate and high background events. In the Belle II experiment, the Level 1 trigger allows reducing the trigger rate to 30kHZ while the beams cross at 40MHz [1]. This L1 trigger aggregates information from several sensors and takes a decision within 5us based on energy and tracks reconstruction. Deploying such fast online decision logic forces to use of dedicated algorithm and computation platform, namely FPGA. Even though the L1 trigger uses data from several sensors, it relies mostly on the Central Drift Chamber (CDC) for its track trigger and makes poor use of the silicon based vertex detector (or inner tracker) [2]. It can be justified by the difficulties both to extract the high data rate from the inner tracker and to extract track information from this high data volume. An upgraded version of the inner tracker planned in the coming years will likely install 5 or 7 layers of CMOS Monolithic Active Pixel Sensor close to the Interaction Point, providing precise information on time and position of particles. This upgrade could (if taken into account during the design phase) provide a more efficient solution to read the information from the inner tracker, making them available for the L1 trigger. If the precise hit information could become available for the L1 trigger, the question of an appropriate data processing method for pixel-based track reconstruction remains open. Several tracking algorithms exist, including some based on Learning approaches. More specifically, Dr B. Schwenker from univ Göttingen developed a Graph Neural Network (GNN) tracking algorithm tailored for this inner tracker upgraded pixel sensor. The algorithm shows good performances in the Belle II Analysis Software Framework (BASF2) but its original implementation is not suited for use in the L1 trigger due to its huge computational requirements. This master thesis presents the efforts to deploy this GNN tracking algorithm on FPGA to try and meet the L1 trigger requirement. It should be compared to similar work in [3, 4] that also target using GNN on FPGA for real-time particle tracking. The main difference from our work is that [3, 4] focus on ATLAS based dataset.

Note: Presented on 15 09 2022
Note: MSc