Real-time object detection for neutrino interactions in the DUNE trigger system

Abstract

The Deep Underground Neutrino Experiment (DUNE) is a next-generation long baseline neutrino experiment with several ambitious physics goals including precision measurements of the neutrino mass hierarchy, leptonic CP-violation in neutrino oscillations, search for proton decay, and the study of neutrino burst signal from core collapse supernovae. The DUNE far detector (FD) will consist of 4 liquid-argon time projection chamber (LArTPC) modules with a total 40 kt of fiducial volume. Data will be read out at the rate of ∼1.2 TB/s per module with the goal of zero down time for the detector. However, a 30 PB per year storage limit is required which demands an accurate, low-latency data selection system.

To detect neutrinos originating from a supernova event, it is imperative to distinguish neutrino interactions in the Ο(1−100) MeV domain from radiogenic backgrounds that dominate the low energy regime. The rise of deep learning (DL) algorithms may have the potential to also improve the triggering system in DUNE by leveraging the unprecedented accuracy of novel DL models and near real-time inference capabilities on modern GPUs. This thesis presents an DL object detection method for low-energy neutrino interaction detection trigger primitive data using the You-Only-Look-Once version 3 (YOLOv3) network.

Extensive studies of optimisation methods surrounding image synthesis and downsampling of simulated LArTPC data are presented in the context of improving radiogenic background rejection and model inference speed. A neutrino detection efficiency of ε_ν = 76.2 ± 0.2% was achieved at a false positive rate of R_FP+ = 1.8^+2.1_−1.8 Hz per FD module for neutrino energies between 1 to 70 MeV. At this false positive rate it is possible to maintain efficiency above 80% for E_ν > 20 MeV. A rapid decrease in detection efficiency to a plateau at approximately 50% is seen for E_ν < 10 MeV. Tuning different control variables enables to select higher detection efficiencies such as ϵ_ν = 85.6 ± 0.2% at R_FP+ = 10.4^+3.9_−3.6 Hz/FD or lower false positive rates like R_FP+ = 0.4^+0.7_−0.4 Hz/FD with ε_ν = 66.7 ± 0.1%.

The performance of a supernova burst trigger using a simple counting method from individual YOLOv3 detections was evaluated using a model of unoscillated supernova neutrino flux and limiting false trigger rate to one per month necessary to meet the DUNE trigger requirements. A trigger efficiency of ε_SNB = 100% was achieved within the range of the Milky Way galaxy while satisfying the technical requirements. For distances reaching the Large Magellanic Cloud (50 kpc), the efficiency reduces to ε_ν ≈ 55%. By incorporating a pre-filtering method based on the overall brightness of images it is shown that the system can process images at a rate of at least 0.768 ms per image and faster using a GPU thus providing sufficient throughput for triggering purposes in DUNE.

Date of Award	1 Oct 2024
Original language	English
Awarding Institution	University of Bristol
Supervisor	Jim Brooke (Supervisor) & Henning U Flaecher (Supervisor)