Photon showers in the ATLAS fast calorimeter simulation: A voxelized dataset with minimized information loss and improved ML models

For the upcoming high luminosity runs at ATLAS, it is necessary to replace the full calorimeter simulation with the fast calorimeter simulation in almost all cases. The current ATLAS fast calorimeter simulation, AtlFast3, uses parametric approaches to speed up the computing time significantly. Even though AtlFast3 generally achieves a high degree of accuracy, further quality improvements within the fast calorimeter simulation are necessary to enable this replacement. In this note, we present a new voxelization scheme for fast calorimeter simulation, with a specific focus on photon-induced showers. This updated discretization significantly reduces several artifacts previously observed in reconstructed photon shower shapes. While this work concentrates on photons, extensions to other particle types are foreseen in future developments. The resulting photon dataset is expected to be released for public usage in the near future. Furthermore, it is shown that the voxelized dataset can be learned by normalizing flows and diffusion models, which are able to generate highly similar datasets afterwards.