Speaker
Description
In this work, we investigate diffusion-based generative models as a fast simulation alternative for modeling detector response on the example of the electromagnetic calorimeter response for the LHCb experiment. We consider both classical denoising diffusion probabilistic models with Gaussian noise and their extension based on Gamma-distributed noise, which is expected to be better suited for positive, skewed energy deposition patterns characteristic of electromagnetic showers.
The proposed U-Net–based architecture with self-attention and conditioning is employed to capture the complex spatial structure of energy showers. The results demonstrate that diffusion models can successfully reproduce both the global and local properties of Geant4-simulated energy showers, achieving high agreement in terms of generative quality and physical consistency. In particular, Gamma-based diffusion models show competitive performance, highlighting the potential benefits of non-Gaussian noise formulations for calorimeter simulation tasks.
In addition, several approaches aimed to accelerate the inference of studied diffusion-based generative models. Our results indicate that diffusion models constitute a promising and scalable solution for fast detector simulations in current and future high energy physics experiments.