Speaker
Description
To address the sharply increasing simulation computing cost faced by the CEPC strip-crystal electromagnetic calorimeter (ECAL) under high-luminosity operating conditions, this work focuses on fast simulation for the CEPC strip-crystal ECAL. We aim to develop a fast simulation and digitization scheme that balances speed and accuracy. High-fidelity training data will be produced by combining Geant4 full simulation with Opticks optical simulation. Using a Diffusion Transformer (DiT), we will investigate Pixel DiT (PiT)–based machine-learning modeling to learn the mapping from voxelized shower energy deposition to detector readout signals, targeting a 10–100× inference speedup while keeping energy and timing residuals within 5%. In addition, a unified inference interface based on ONNX will be built and integrated into the CEPCSW software framework, and the overall impact on reconstruction and physics analysis will be evaluated.