Generative Artificial Intelligence for Air Shower Simulation

TL;DR

This work tackles the computational bottleneck of detailed extensive air shower simulations by introducing GAIAS2, a generative surrogate based on a Wasserstein GAN with Gradient Penalty and Self-Attention trained on CORSIKA proton-induced showers. A fixed 4D tensor representation and data-driven binning enable efficient learning of ground-level muon distributions, while an ensemble of 57 generators mitigates mode-coverage issues and achieves a Wasserstein distance of about 0.04 relative to training data. The trained model produces tens of thousands of samples in under a minute on a single GPU, offering roughly a $10^4$x speed-up over full MC simulations and a pathway toward real-time conditioning and broader applicability in astroparticle experiments. This approach is detector-agnostic and can be extended to multiple particle species and integrated with differentiable or hybrid simulation pipelines for rapid, scalable cosmic-ray studies.

Abstract

The detailed simulation of extensive air showers, produced by primary cosmic rays interacting in the atmosphere, is a task that is traditionally undertaken by means of Monte Carlo methods. These processes are computationally intensive, accounting for a major fraction of the computational resources used in the large-scale simulations required by current and future experiments in the field of astroparticle physics. In this work, we present a novel approach based on Generative Adversarial Networks (GANs) to accelerate air shower simulations. We developed and trained a GAN on a dataset of high-energy proton-induced air showers generated with \texttt{CORSIKA}; our model reproduces key distributions of secondary particles, such as energy spectra and spatial distributions at ground level of muons. Once the model has been trained, which takes approximately 74 hours, the generation real time per shower is reduced by a factor of $10^4$ with respect to the full \texttt{CORSIKA} simulation, leading to a substantial decrease in both computational time and energy consumption.

Figures (14)

• Figure 1: Processing time per event (real time, as from the Unix command time) for benchmark simulations in the four different servers used in this work for the Monte Carlo generation of the training samples. The different colours reported in the figures account for the different hadronic interaction models used in the simulations (see text for more details).
• Figure 2: Azimuthal angle ($\phi$) distribution (top) and spatial distribution of muon positions in the $(x, y)$ plane (bottom) at ground level.
• Figure 3: Ratio of negative to positive muons ($\mu^-/\mu^+$) at ground level as a function of momentum.
• Figure 4: Muon multiplicity distributions at ground level (top) and normalised energy spectra (bottom) for air showers induced by primary protons of different energies.
• Figure 5: Longitudinal profile of the air shower. The vertical red dotted line marks the atmospheric depth corresponding to the shower maximum.