17–31 Jul 2025
Orthodox Academy of Crete, Kolymbari, Crete, Greece
Europe/Athens timezone
Please see new information on proceedings at the link "Scientific Information"-> "Proceedings"

Machine Learning–Based Real-Time Data Processing and Hardware Deployment in JUNO experiment

28 Jul 2025, 09:00
30m
Room 1

Room 1

Talk Special session on Machine Learning Special Session on Machine Learning

Speaker

Feng Gao (iihe, ULB)

Description

In current-generation particle detectors, traditional backend data processing is increasingly migrating toward front-end electronics. This shift enables earlier event selection and real-time signal processing within the data acquisition chain, reducing bandwidth and improving system responsiveness. In large-scale neutrino experiments, identifying low-energy events is particularly difficult due to weak signal amplitudes and overlap with dark noise, requiring early-stage processing systems with effective background suppression and resource-efficient implementation. This work investigates how machine learning methods can be adapted to address potential future requirements on latency, power efficiency, and hardware compatibility in this context.

This work investigates the feasibility of applying machine learning techniques to enhance the online processing of low-energy events—not only for L1 trigger decisions but also for signal identification at the data acquisition (DAQ) and preprocessing stages. We present a comprehensive comparison of three deployment strategies: high-level synthesis using hls4ml on FPGAs, manual RTL implementation (Verilog), and inference execution on Deep Learning Processing Units (DPUs). These approaches are evaluated in terms of resource usage, response latency, and system integration trade-offs.

Training datasets were constructed using simulated events from the Jiangmen Underground Neutrino Observatory (JUNO) SNiPER framework, including both low-energy electron signals and PMT dark noise. Models were trained on a conventional CPU-based system and deployed on Jetson Nano, Kintex-7 FPGA, and Virtex-7 FPGA hardware for evaluation, assessing their adaptability and performance across platforms with varying computational capacity.

This study proposes a machine-learning-driven data processing framework tailored to future neutrino experiments, providing practical deployment insights for building high-performance, low-power, and scalable real-time systems.

Details

Feng Gao, Postdoctoral Researcher, Interuniversity Institute for High Energies, Université Libre de Bruxelles, Belgium

Internet talk No
Is this an abstract from experimental collaboration? Yes
Name of experiment and experimental site Jiangmen Underground Neutrino Observatory(JUNO experiment), Jiangmen, China
Is the speaker for that presentation defined? Yes

Author

Feng Gao (iihe, ULB)

Co-authors

Barbara Clerbaux (Universite Libre de Bruxelles (BE)) yifan yang (iihe)

Presentation materials