Graph Neural Networks for Software Compensation and π⁰/γ Separation in a High-Granularity Calorimeter

We present a design for a high-granularity zero-degree calorimeter (ZDC) tailored for the upcoming Electron-Ion Collider (EIC), using SiPM-on-tile technology. The design features a novel hexagonal staggered layer arrangement that improves spatial resolution. To fully leverage the design's high granularity, we employ graph neural networks (GNNs) for energy and angle regression as well as signal classification. The GNN approach yields a compensated response with energy and angular resolution of 34%/√E and 0.66 mrad/√E respectively for single neutron showers, thus meeting the requirements set in the EIC Yellow Report. Additionally, the GNN leads to 98% efficiency in classifying π⁰ versus γ showers at energies above 150 GeV. Our studies also demonstrate that graph neural networks can substantially enhance the performance of high-granularity CALICE-style calorimeters by automating and optimizing the software compensation algorithms required for these systems.