Source-linked AI summary
Precise simulation of electromagnetic calorimeter showers using a Wasserstein Generative Adversarial Network
Martin Erdmann, Jonas Glombitza, Thorben Quast
TL;DR
Detailed calorimeter shower simulations are computationally expensive because they must reproduce stochastic, high-dimensional energy depositions and correlations. The paper trains a Wasserstein GAN on GEANT4 showers, using constrainer networks to preserve electron energy and impact-position dependencies. The generated showers reproduce most evaluated observables and correlations at a level comparable to GEANT4, with a several-orders-of-magnitude speed-up but underestimated low-energy pixel densities.
Problem
Sequential calorimeter-shower simulations require substantial computing resources, while realistic generation must capture high-dimensional depositions, fluctuations, correlations, and beam-condition dependencies.
Method
The paper trains a Wasserstein GAN with a generator, critic, and supervised constrainer networks that enforce electron-energy and impact-position dependencies using GEANT4 reference showers.
Results
In most aspects, WGAN showers reach GEANT4 quality, reproducing typical observables and correlations; the study reports several orders of magnitude computational speed-up.
Takeaways & Limitations
WGANs can successfully simulate isolated electromagnetic showers in a realistic multilayer sampling calorimeter while preserving key shower properties and correlations.
Takeaways & Limitations
The WGAN underestimates low-energy-density pixels; restricting analysis to pixels above 10 MIP equivalents excludes a component contributing 10% of the total signal.
Abstract
from arXiv · showhide
Simulations of particle showers in calorimeters are computationally time-consuming, as they have to reproduce both energy depositions and their considerable fluctuations. A new approach to ultra-fast simulations are generative models where all calorimeter energy depositions are generated simultaneously. We use GEANT4 simulations of an electron beam impinging on a multi-layer electromagnetic calorimeter for adversarial training of a generator network and a critic network guided by the Wasserstein distance. The generator is constraint during the training such that the generated showers show the expected dependency on the initial energy and the impact position. It produces realistic calorimeter energy depositions, fluctuations and correlations which we demonstrate in distributions of typical calorimeter observables. In most aspects, we observe that generated calorimeter showers reach the level of showers as simulated with the GEANT4 program.
1 Introduction
Calorimeter shower simulation is computationally demanding because realistic showers involve stochastic, high-dimensional energy depositions and correlations. The paper applies a Wasserstein GAN with conditioning networks to generate complete electromagnetic showers while respecting initial beam conditions.
- GEANT4 accurately models four-dimensional shower development but sequentially simulating individual interactions requires substantial computing resources.
- Generative models instead produce the shower’s three-dimensional spatial energy distribution directly, automatically encoding a high-dimensional deposition probability distribution.
- The central challenge is training a generator that approximates the full high-dimensional probability distribution of realistic calorimeter energy depositions.
- A generator can determine all spatial energy depositions in one evaluation, using random inputs to produce distinct stochastic shower realizations.
- The paper replaces binary GAN evaluation with Wasserstein-distance training because conventional calorimeter GAN training was only partially successful.
- The conditioned WGAN targets showers for specified impact positions and electron energies in a realistic multilayer pixelated electromagnetic calorimeter.
2 Experimental setup
The study uses a seven-layer electromagnetic sampling calorimeter prototype and GEANT4-generated electron showers as reference data. Training covers five beam energies, while a separate 70 GeV sample tests interpolation to an unseen energy.
- The selected HGCAL electromagnetic-compartment prototype has seven sensitive layers spanning 2.8–16.2 radiation lengths.
- Each calorimeter sensor is a 6-inch hexagonal silicon n-type wafer with 300 µm active thickness at full depletion.
- GEANT4 version 10.2 supplies the reference electromagnetic-shower dataset used to train and evaluate the WGAN.
- The beam setup models perpendicular electron incidence, track-based impact positions, a 6×5 cm2 rectangular profile, and 1% energy smearing.
- Pixels with depositions below 2 MIPs are removed to reject noise contributions.
- Training uses 5 × 100,000 showers from 20, 32, 50, 80, and 90 GeV electrons, while simulated 70 GeV electrons test interpolation.
3 Fast simulation approach
The approach combines a Wasserstein GAN with label conditioning and supervised constrainer networks to generate calorimeter showers while respecting electron energy and impact position. Generated showers are evaluated against GEANT4 simulations through energy depositions, occupancy, and reconstructed labels.
- Wasserstein GAN: The GAN framework trains a generator against a critic so generated shower samples approximate realistic calorimeter data.The Wasserstein formulation replaces binary discrimination with a distance-based critic objective and gradient penalty.
- Label conditioning: The generator receives noise, electron energy, and impact position, while the critic also receives the physics labels.The generator is modified to G(z, E, P), with P = (Px, Py).
- Label conditioning: Two supervised constrainer networks reconstruct energy and impact position from real showers and guide label reconstruction for generated showers.The generator minimizes both the Wasserstein distance and an auxiliary loss based on the constrainer outputs.
- Network framework: The framework comprises one generator, one critic, and two constrainer networks for conditioning energy and impact position.The networks are implemented and evaluated as a coordinated framework for electromagnetic shower generation.
4 Performance benchmarks
The WGAN generates electromagnetic showers whose observables and correlations generally agree with GEANT4, including for 70 GeV showers excluded from training, while reducing simulation time by up to three orders of magnitude. Its main mismatch is an underrepresentation of low-energy pixel depositions, which affects active-pixel counts and energy sums.
- Visual inspection: WGAN-generated showers reproduce expected energy scaling, impact-position dependence, and layerwise shower structure in visual inspections.Pixel occupancies and intensities scale with incident energy, largest depositions shift with impact position, and activity peaks in intermediate layers.
- Visual inspection: The radial pixel-occupancy profile resembles GEANT4, although the overall WGAN occupancy scale is underestimated.The comparison uses 90 GeV showers and a 2 MIP threshold.
- Distributions of calorimeter observables: WGAN and GEANT4 agree reasonably for total energy deposition, longitudinal shower depth, per-layer maximum pixel energy, and transverse shower spread.The transverse-shape agreement is representative across layers and coordinates except for the first layer at 2.8 X0.
- Distributions of calorimeter observables: The WGAN underrepresents pixel depositions below ≈10 MIPs, causing mismatches in active-pixel counts and total energy sums.Restricting analyses to pixels above 10 MIP equivalents isolates the well-described range; the rejected spectrum contributes 10% of the total signal in this setup.
- Correlations: Generated showers reproduce GEANT4 pairwise correlations for several observables, but the correlation between active-pixel count and energy sum is not matched.The latter discrepancy is attributed in the passage to the underestimated low-energy spectrum.
- Computational speed-up: WGAN evaluation is up to three orders of magnitude faster than full GEANT4 simulation, with evaluation time independent of incident electron energy.The reported speed advantage is compared with expert-engineered parameterizations providing CPU gains of 10-100 over GEANT4.
5 Conclusion
The WGAN successfully generates isolated electromagnetic showers in a realistic multilayer calorimeter, with simulation speed-up of several orders of magnitude and quality reaching GEANT4 levels in most aspects.
- 5 Conclusion: The WGAN combines a generator, critic, and constrainer networks to model showers while preserving dependencies on primary energy and impact location.The critic approximates Wasserstein distance during training, while constrainer networks enforce the desired energy and position dependencies.
- 5 Conclusion: WGAN-generated showers reproduce typical high fluctuations, sparse energy depositions, and average pixel occupancy comparable to GEANT4 showers.These properties were observed through visual inspection and average pixel-occupancy comparisons.
- 5 Conclusion: Longitudinal depths, layer energy depositions, maximum sensor energy, and transverse shower shapes generally agree well between WGAN and GEANT4 showers.Longitudinal shower shapes were reproduced even for an intermediate electron energy absent from training.
- 5 Conclusion: The WGAN underestimates low-energy-density pixels, reducing the number above the 2 MIPs threshold by about 15%.This is the principal reported discrepancy among the inspected calorimeter observables.
- 5 Conclusion: Energy-deposition correlations between neighboring layers and correlations between total energy and pixel counts show strong agreement with expected behavior.The longitudinal shower depth decreases with increasing total energy sum for this sampling configuration.
- 5 Conclusion: The study covers isolated showers, while extensions to full collision events require additional background labels or assumptions about electronic and saturation effects.Independent simultaneous deposits could be superimposed only if electronic and saturation effects can be neglected.
- 5 Conclusion: Compared with traditional sequential simulations, the WGAN provides a speed-up of several orders of magnitude while reaching GEANT4 shower quality in most aspects.The conclusion characterizes this as ultra-fast simulation of isolated electromagnetic showers in a realistic multilayer sampling calorimeter.
A Appendix
The appendix documents the critic, generator, and constrainer networks, along with figures reproduced under a higher noise threshold.
- A Appendix: Table 2 describes the critic network used in the adversarial framework.
- A Appendix: Table 3 describes the generator network used to generate electromagnetic calorimeter showers.
- A Appendix: Table 4 describes the constrainer network used for energy and position regression.
- A Appendix: Figures 10 and 11 reproduce earlier figure analyses with the noise threshold increased from 2 MIPs to 10 MIPs.