Neural Network for Simulating Radio Emission from Extensive Air Showers

TL;DR

This work tackles the computational bottleneck of simulating radio emission from extensive air showers by introducing a neural-network surrogate that deterministically predicts antenna pulses across diverse geometries. The model ingests physically meaningful inputs, including $X_{max}$ and $E_{em}$, and outputs time-resolved pulses for both polarizations at multiple antenna positions, enabling rapid $X_{max}$ reconstruction. Results show pulse predictions qualitatively match CoREAS, with fluence agreement around 10% and comparable bias and resolution in $X_{max}$ reconstruction relative to full Monte-Carlo simulations; differences between emission models remain within detector uncertainties, suggesting broad applicability. The approach offers substantial speedups, low memory footprint, and differentiability, making it a practical surrogate for large radio arrays and a flexible starting point for fine-tuning to other experiments or extended inputs such as longitudinal shower profiles.

Abstract

Cosmic ray shower detection using large radio arrays has gained significant traction in recent years. With massive improvements in signal modelling and microscopic simulations, the analysis of incoming events is still severely limited by the simulation cost of radio emission to interpret the data. In this work, we show that a neural network can be used for simulating such radio pulses. This work serves as a proof of concept that simple neural networks can be used for emergent deterministic macroscopic phenomena of microscopic simulations. We also demonstrate how such a neural network can be used for the physics use case of $X_\mathrm{max}$ reconstruction, while retaining comparable resolution to using full Monte-Carlo simulations for radio emission. Code available at https://anonymous.4open.science/r/radio_nn-21BF/.

Figures (3)

• Figure 1: Neural network schematic for generating radio pulses: The 11 inputs are described in \ref{['sec:nn_model']}, and the 512 outputs are the 256 time bins in the 2 polarisations. The network produces a pulse, given an antenna position and shower parameters. The model has about 4M parameters, and has a memory footprint of 19MB.
• Figure 2: This figure on the left shows the change in loss functions as a function of training steps (or epochs). The figure on the right shows how the learning rate is varied with the training steps. In the plot on the left, the typical deviation from the true $X_\mathrm{max}$ for a $X_\mathrm{max}$ reconstruction procedure for a random shower is shown.
• Figure 3: Comparison of generated radio pulses between a CoREAS simulation and neural network-based simulation for both polarizations at antennas at various distances from the shower core (R). This is a simulation of a single event.