Speaker
Description
The Electron–Ion Collider (EIC) will generate continuous, high-rate detector streams, creating a need for low-latency identification of rare physics signals amid substantial background and pileup. We propose a streaming-native, physics-informed AI framework that brings inference directly into the detector data-processing pipeline. Detector readout is organized into time-ordered frames that preserve temporal correlations and detector response without requiring predefined event boundaries. The architecture combines selective state-space models for efficient sequence processing, joint-embedding predictive learning for self-supervised representation learning, and Streaming Latent Event Representation Modeling (SLERM) to capture relationships among detector signals without explicit graph construction. Energy–momentum conservation, kinematic consistency, and detector-response priors are incorporated into the training objective to promote physically consistent predictions. Using the Sullivan process as a benchmark, the framework will be evaluated on simulated signal and background streams and pseudo-streaming data. Planned integration into EIC workflow will support comparison with conventional reconstruction and filtering workflows through signal efficiency, background rejection, reconstruction fidelity, inference latency, and sustained throughput. This work aims to reduce the time from detector readout to physics insight while preserving sensitivity to rare processes in EIC-scale streaming environments.