CaloClouds3: Ultra-Fast Geometry-Independent Highly-Granular Calorimeter Simulation

TL;DR

CaloClouds3 addresses the challenge of fast, geometry-agnostic photon shower simulation in high-granularity calorimeters by introducing angular conditioning and a pointcloud-based architecture that combines a normalising flow (ShowerFlow) with a distilled diffusion model. Trained on a position-agnostic, regularised ILD SiECAL barrel dataset and integrated into a full simulation/reconstruction chain, it maintains fidelity to Geant4 observables (energy, occupancy, and angular distributions) while achieving substantial inference speedups. The work demonstrates effective angular generalisation, improved reconstruction-relevant metrics (including internal angle with a small subset of hits), and competitive di-photon separation, supporting the potential replacement of Geant4 for fast simulation in practical workflows. It also highlights the benefits of geometry-aware data representations and re-evaluates hyperparameter choices when transferring ML methods from natural-image domains to physics simulations.

Abstract

We present CaloClouds3, a model for the fast simulation of photon showers in the barrel of a high granularity detector. This iteration demonstrates for the first time how a pointcloud model can employ angular conditioning to replicate photons at all incident angles. Showers produced by this model can be used across the whole detector barrel, due to specially produced position agnostic training data. With this flexibility, the model is usable in a full simulation and reconstruction chain, which offers a further handle for evaluating physics performance of the model. As inference time is a crucial consideration for a generative model, the pre-processing and hyperparameters are aggressively optimised, achieving a speed up factor of two orders of magnitude over Geant4 at inference.

Figures (19)

• Figure 1: Photon showers, with varying incident angle, in the top barrel segment of the ILD SiECAL as simulated by Geant4. The incident energy of the photon gun is indicated beneath each shower.
• Figure 2: Differences observed while training the ShowerFlow model for CaloClouds2 and CaloClouds3. In the top left, we observe that the training loss of CaloClouds2 never plateaus. In the top right, we see that the gradient normal of CaloClouds3 stabilised during the training, but the gradient normal of CaloClouds2 continues to increase. The relative magnitudes of the mean and maximum parameters in CaloClouds3 remain in fairly constant relationship.
• Figure 3: The model architecture of CaloClouds3. Parts belonging to ShowerFlow are in purple, and those belonging to the diffusion model are in green. A full description is provided in section \ref{['sec:CC3arch']}.
• Figure 4: Energy distributions of CaloClouds2 (CC2) and CaloClouds3 (CC3), in comparison to Geant4. Top row are the distributions themselves, averaged over three different seeds, with error bars from the standard deviation. Bottom row are ratio plots between each model and Geant4, with errors propagated from the distributions. Left, histogram of the energy of each cell, with the MIP cut marked as a hatched box. Centre, radial profile of the energy. Right, distribution of the energy through the layers.
• Figure 5: Occupancy distributions of CaloClouds2 (CC2) and CaloClouds3 (CC3), in comparison to Geant4. Top row are the distributions themselves, averaged over three different seeds, with error bars defined by the standard deviation. Bottom row are ratio plots between each model and Geant4, with errors propagated from the distributions. Left, histogram of the occupancy of the calorimeter in each shower. Centre, the number of active cells in concentric rings about the shower axis. Right, the number of active cells in each layer.