Speakers
Description
Dual-readout calorimetry achieves superior energy resolution by simultaneously measuring Cherenkov and scintillation light for event-by-event corrections, making it attractive for future Higgs factories. However, extracting these components via full waveform digitization poses significant data-rate challenges for front-end readout. We present a systematic comparison of machine learning (ML) and template fitting approaches for real-time separation of Cherenkov and scintillation signals in homogeneous dual-readout calorimeters, evaluated across three representative crystal types (BGO, BSO, PWO). ML models match or exceed template fitting performance at substantially lower sampling rates, and a single model trained across multiple incident energies generalizes robustly without retraining. FPGA-compatible compression via pruning and quantization further enables latencies suitable for real-time deployment. We also assess the downstream impact of this approach on particle reconstruction performance, offering an early indication of its viability for future detector design and readout implementation.
