SRO AIAI for Streaming Readout

Mission ​

The Electron-Ion Collider will read out its detectors continuously at rates near 100 Gb/s. Rare physics signals have to be found in real time, inside unsegmented data with heavy background and pileup. Trigger and reconstruction frameworks built for discrete, low-rate events do not scale to this regime, and offline machine learning pipelines add latency that limits sensitivity to rare processes.

SRO AI develops stream-native, physics-integrated AI that runs at the data acquisition layer: continuous signal–background discrimination, event prioritization and anomaly detection as the data are produced.

Research thrusts ​

• Data layer — a unified streaming representation for Geant4 simulation and Jefferson Lab pseudo-streaming data: synchronized detector channels, normalization across run conditions, sliding temporal windows.
• AI layer — a state-space backbone, JEPA-style self-supervised learning and SLERM, with physics constraints in the loss.
• Benchmarking — signal efficiency, background rejection, reconstructed kinematics, latency and throughput, measured against event-based pipelines as data rate and pileup grow.
• Integration — real-time inference inside the JANA2 framework, with model compression and quantization for streaming deployment.

The physics benchmark is the Sullivan process — scattering off the pion cloud of the nucleon, a key channel for meson structure at the EIC — studied against inclusive DIS and meson production backgrounds at high luminosity.

Projects ​

Institutions ​

• The Catholic University of America
• Thomas Jefferson National Accelerator Facility