Description
At the Electron-Ion Collider, the reliable identification of physics processes of interest demands robust strategies for suppressing instrumental backgrounds and for resolving distinct reaction types within complex, high-multiplicity event environments.In the context of streaming readout architectures, data-driven machine learning approaches offer an automated pathway to event selection and physics discrimination.
We present a machine learning framework that operates at two complementary levels of detector granularity: anomaly identification within tracker hit collections across multiple detector subsystems, and classification of physics reaction types at the event level using calorimeter and tracking observables. The framework ingests reconstructed detector responses across a hierarchy of observables - from low-level geometric and kinematic track parameters to higher-order event-level quantities while constructing feature representations informed by the underlying physics and tailored to distinguish signal-like reactions from process-level backgrounds. The choice of observables is motivated by generalizability both across heterogeneous detector technologies comprising the ePIC detector and across the range of physics processes accessible at the EIC. Building upon our prior work employing complementary unsupervised methods for hit-level anomaly characterization, we introduce a classification approach for event-level discrimination between inclusive deep inelastic scattering and process-specific reactions, trained on mixed datasets. The approach targets a hierarchy of discrimination objectives: identification of noise at the sub-event level, and the dominant physics reaction at the event level, yielding interpretable outputs for both background characterization and physics process tagging.
The present study constitutes a proof-of-concept demonstration, and lays the groundwork for a scalable, multi-method framework for automated physics discrimination at the EIC, with the present prototype intended as a foundation for expansion toward a unified ensemble methodology as datasets and detector commissioning mature.
