CaloMan: Fast generation of calorimeter showers with density estimation on learned manifolds

Jesse C. Cresswell, Brendan Leigh Ross, Gabriel Loaiza-Ganem, Humberto Reyes-González, Marco Letizia, Anthony L. Caterini

Introduction

In high-energy physics, researchers at the Large Hadron Collider (LHC) rely on massive amounts of simulated data to interpret experimental results. The current gold standard for simulation, Geant4, provides exceptional accuracy but is computationally prohibitive, often requiring seconds or minutes to simulate a single particle "shower"—the cascade of secondary particles produced when a high-energy particle strikes a detector.

CaloMan is a deep generative model designed to address this bottleneck. By leveraging the "manifold hypothesis"—the principle that high-dimensional data often resides on a much simpler, low-dimensional surface—CaloMan achieves generation speeds as fast as 0.02 ms per event, outperforming existing surrogate models while maintaining high physical fidelity.

Methodology: Respecting the Data's Natural Dimensionality

Standard Deep Generative Models (DGMs) like GANs and Normalizing Flows often suffer from "manifold overfitting." Calorimeter data is high-dimensional (comprising hundreds or thousands of voxels), but the physical laws of Quantum Electrodynamics (QED) constrain this data to a thin, low-dimensional slice of the total space. Attempting to model the entire high-dimensional volume leads to inaccuracies.

CaloMan employs a two-step framework to respect these constraints:

  1. Manifold Learning: A Variational Autoencoder (VAE) compresses the 368-dimensional voxel data into a compact latent space.
  2. Density Estimation: A Normalizing Flow (specifically a Rational-Quadratic Neural Spline Flow) models the probability distribution within this low-dimensional latent space, conditioned on the particle's incident energy.
Estimated intrinsic dimension of the photon dataset using the Levina-Bickel estimator, showing that the 368-voxel showers reside on a manifold of approximately 20 dimensions.

A critical innovation in this work is the use of the Levina-Bickel estimator to scientifically determine the data's "intrinsic dimension." The researchers found that the 368-voxel showers actually live on a manifold of approximately 20 dimensions. By reducing dimensionality first, the subsequent Normalizing Flow avoids learning noise and becomes significantly faster to sample.

Performance and Results

The model was evaluated using the Fast Calorimeter Simulation Challenge 2022 (Dataset 2), which contains 121,000 electromagnetic showers simulated via Geant4.

Speed and Fidelity

CaloMan demonstrates remarkable efficiency, generating showers in 0.02 ms per event on a single GPU. For comparison, the competing CaloScore model requires 40 ms, and the original Geant4 simulation can take up to 3,000 ms.

To assess realism, a binary classifier was trained to distinguish between real and generated showers. CaloMan achieved an Area Under the Curve (AUC) of 0.78. While a perfect score of 0.5 indicates indistinguishable samples, CaloMan significantly outperformed CaloScore, which recorded an AUC of 0.98.

Comparison of average energy deposition across five detector layers between the Geant4 test dataset and CaloMan generated samples, demonstrating high physical fidelity.

Physics Accuracy

High-level physical features, such as the center of energy and energy deposition per layer, show strong agreement with the Geant4 reference data. This confirms that the manifold-based approach successfully captures the underlying physics of particle interactions without the computational overhead of traditional solvers.

Conclusion and Future Work

CaloMan serves as a proof of concept that manifold learning can drastically accelerate physics simulations. Future work will focus on scaling this method to more complex geometries, such as the 45,000-voxel datasets found in advanced detector designs. Additionally, the authors suggest that conditioning the manifold learning step itself on incident energy could further improve the model's ability to generalize across unseen energy ranges.

Summary updated

Abstract

Precision measurements and new physics searches at the Large Hadron Collider require efficient simulations of particle propagation and interactions within the detectors. The most computationally expensive simulations involve calorimeter showers. Advances in deep generative modelling - particularly in the realm of high-dimensional data - have opened the possibility of generating realistic calorimeter showers orders of magnitude more quickly than physics-based simulation. However, the high-dimensional representation of showers belies the relative simplicity and structure of the underlying physical laws. This phenomenon is yet another example of the manifold hypothesis from machine learning, which states that high-dimensional data is supported on low-dimensional manifolds. We thus propose modelling calorimeter showers first by learning their manifold structure, and then estimating the density of data across this manifold. Learning manifold structure reduces the dimensionality of the data, which enables fast training and generation when compared with competing methods.
BibTeX
@misc{cresswell2022caloman,
  author = {Jesse C. Cresswell and Brendan Leigh Ross and Gabriel Loaiza-Ganem and Humberto Reyes-González and Marco Letizia and Anthony L. Caterini},
  title = {{CaloMan: Fast generation of calorimeter showers with density estimation on learned manifolds}},
  year = {2022},
  eprint = {2211.15380},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2211.15380},
}
Figure