Description
Future experiments at Jefferson Lab (JLab) and the Electron-Ion Collider (EIC) are designed to leverage the highest achievable luminosities to probe hadronic structure with unprecedented precision. These measurements require the efficient identification of rare physics events in the presence of substantial backgrounds. In this work, measurements of meson structure functions through the Sullivan process serve as the physics-validation case for developing and evaluating AI/ML processing within a streaming-readout environment.
To validate the approach, existing data from Jefferson Lab (Jlab) acquired with the Neutral Particle Spectrometer, a PWO based electromagnetic calorimeter were used. Transformer-based and graph neural network (GNN) models were trained for cluster reconstruction and anomaly detection using existing meson data. These models have been integrated into the JANA2 reconstruction framework which enables GPU-accelerated inference. In parallel, a separate neural-network model has been implemented on an FPGA to filter background events. This completed work establishes the heterogeneous CPU, GPU, and FPGA processing chain needed for real-time reconstruction and event selection.
The next step will be a rate capability test using a beam test at JLab using the small-scale NPS prototype electro-magnetic calorimeter. JLab FADC250 and VXS Trigger Processor (VTP) modules will provide the bandwidth required for streaming readout. This beam test will provide an
estimation of achievable throughput and latency for the streaming-readout system while evaluating the applicability of the AI/ML-based models under realistic experimental conditions.
