CaloGraph - Fast Simulation for Detector Calorimeters

Simulating energy deposition in calorimeters is a fundamental component of event generation in high-energy physics. These simulations are indispensable for developing reconstruction algorithms, physics analyses and phenomenological studies. Accurate modelling of calorimeter responses is essential for understanding detector performance and interpreting experimental data.

However, full simulation tools such as Geant4, while highly detailed and accurate, are computationally expensive. This becomes a major limitation when large datasets are required. As experiments move toward higher luminosities and increasingly complex detector environments, the demand for fast, high-fidelity simulation grows rapidly.

CaloGraph [1] addresses this challenge by introducing a generative model tailored to calorimeter simulation. Unlike previous approaches that model showers using images or point clouds, CaloGraph represents the calorimeter as a graph, where each node corresponds to a detector cell and its energy deposit. This graph-based representation is particularly well-suited to irregular or sparse detector geometries and reduces pre- and postprocessing of the data.

At the core of CaloGraph is a denoising diffusion model—a class of generative models that have recently shown state-of-the-art performance in domains such as image and molecule generation. In this framework, the model learns to reverse a stochastic process that gradually perturbs real showers into noise, generating new, physically plausible showers conditioned on the incoming particle’s features.

This new and unique combination of diffusion modelling and graph-based geometry awareness enables CaloGraph to generate realistic showers in complex detector configurations. CaloGraph has been evaluated on ATLAS-like calorimeter datasets and shows promising results in reproducing key shower characteristics.

[1] Dmitrii Kobylianskii et al 2024 Phys. Rev. D 110, 072003