Machine Learning for Event Reconstruction at the Electron-Ion Collider

We present a novel two‐stage particle‐identification (PID) workflow for the ePIC Barrel Imaging Calorimeter (BIC) at the future Electron–Ion Collider. In the first stage, we exploit the classical calorimeter‐to‐track energy ratio E/p (optimally summing energy across the first eight SciFi layers) to achieve a 97 % electron efficiency and a pion rejection factor R_π≈23.5. In the second stage, we reshape high‐granularity per‐hit data into a pseudo-image (layers × top-hits) × feature-channels, encoding normalized hit energy, radial coordinate, angular separations, and sub-detector flags. A VGG-style convolutional neural network trained on these five channels learns residual shower‐shape differences between electrons and pions. By selecting the CNN output threshold to yield an additional 97 % efficiency on the pre-cut sample, we demonstrate an overall electron efficiency of 95 % with a combined pion rejection R_π≈174 with E_beam = 1 GeV, representing nearly an order-of-magnitude improvement over E/p alone. Our work delivers a production-ready ML‐augmented PID module, poised for integration into the EICrecon framework and deployment in upcoming ePIC physics analyses.