Neural Network for Simulating Radio Emission from Extensive Air Showers¶

Conference: NeurIPS 2025 arXiv: 2512.21407 Code: Available Area: AI for Science / Astroparticle Physics Keywords: Cosmic rays, extensive air showers, radio emission simulation, neural network surrogate model, \(X_{\text{max}}\) reconstruction

TL;DR¶

A simple fully connected neural network is employed to replace computationally expensive CoREAS Monte Carlo simulations, enabling fast prediction of radio pulses from extensive air showers (EAS) while achieving \(X_{\text{max}}\) reconstruction resolution comparable to conventional simulations.

Background & Motivation¶

When cosmic rays arrive at Earth from distant astrophysical sources and interact with the atmosphere, they produce extensive air showers (EAS)—cascades of secondary particles. Detecting the radio emission of EAS using antenna arrays is an important approach for studying the energy spectrum and mass composition of cosmic rays.

Core bottleneck: Current reconstruction methods require microscopic simulation for each observed event, computing the electromagnetic radiation contribution of every particle track and superimposing them onto each antenna. Simulating a single event can take several weeks, and the analysis pipeline requires multiple iterative simulations per event to quantify uncertainties.

Existing acceleration methods (interpolation, template synthesis) remain computationally demanding and cannot generalize across different event geometries.

Key observation: Although the microscopic particle interactions within a shower are stochastic, the macroscopic radio emission is deterministic—given shower parameters (\(X_{\text{max}}\), electromagnetic energy, arrival direction, etc.), the pulse shape received at each antenna is fully determined. This reduces the problem from generative (learning a distribution) to regression (learning a deterministic mapping).

Method¶

Overall Architecture¶

A point-cloud-inspired approach is adopted: the entire radio emission prediction is decomposed into per-antenna-location prediction tasks. The network takes shower parameters and antenna position as input, and outputs the radio pulse waveform at that antenna in two polarization directions.

Key Designs¶

Input features (11-dimensional):
- Shower physical parameters: \(X_{\text{max}}\) (atmospheric depth of shower maximum), electromagnetic energy \(E_{em}\), geomagnetic field angle, density and altitude at \(X_{\text{max}}\), primary cosmic-ray energy, zenith angle, azimuth angle
- Antenna position: coordinates in the shower frame (3-dimensional)
Output (512-dimensional):
- 256 time bins per polarization direction (\(v \times B\) and \(v \times v \times B\))
- Time resolution: 1 ns
Network architecture:
- 8-layer fully connected network with approximately 4 million parameters; memory footprint of only 19 MB
- No skip connections: at this depth, vanishing gradients are not observed
- Inputs are normalized to the range \([-1, 1]\)

Loss & Training¶

L1 loss \(\mathcal{L} = |x - y|\) rather than L2—ensuring that weak pulses (which contribute little to L2 loss) are also learned adequately
The second polarization (weaker) is assigned higher weight to ensure both polarization directions are correctly learned
Adam optimizer with weight decay regularization
Learning rate schedule: warmup followed by gradual decay (Figure 2, right)

Training data: The CoREAS simulation library from the Pierre Auger Collaboration's AERA array, comprising 2,158 distinct shower parameter configurations × 27 realizations ≈ 58k simulations, each containing 240 antenna positions. An 80/20 train/test split is applied.

Training cost: Completed within one week on a standard desktop CPU (AMD Ryzen 7 PRO 3700); inference requires only a few hundred milliseconds per antenna position.

Key Experimental Results¶

Pulse Quality Evaluation¶

Metric	Result
Energy flux (fluence) accuracy	Within 10% for the majority of cases
Normalized correlation for strong pulses	Highly correlated
Error for weak pulses (< 1e-2 eV/m²)	Up to 100%, but negligible as they are buried in noise

The paper notes that the discrepancy between the two leading radio emission models (CoREAS vs. ZHS) is itself approximately 10%, placing the neural network's errors within the range of model systematic uncertainties.

\(X_{\text{max}}\) Reconstruction — Bias and Resolution Comparison (Table 1)¶

Simulation	5% noise — Bias	5% noise — Resolution	10% noise — Bias	10% noise — Resolution
CoREAS	-12.87 g/cm²	32.22 g/cm²	-10.36 g/cm²	31.42 g/cm²
Neural Network	-12.71 g/cm²	33.13 g/cm²	-11.79 g/cm²	32.38 g/cm²

Bias and resolution are nearly identical across both noise levels
The differences are far smaller than the systematic uncertainty between radio emission models (~11 g/cm²)

Computational Efficiency Comparison¶

Aspect	CoREAS	Neural Network
Single-event simulation time	Hours to weeks	Milliseconds (per antenna)
Parameter count	—	~4 million
Model size	—	19 MB
GPU parallelization	Difficult	Natively supported

Key Findings¶

A simple fully connected network can capture deterministic macroscopic radio emission without the need for complex generative models
\(X_{\text{max}}\) reconstruction performance is comparable to full CoREAS simulations, validating the feasibility of the neural network surrogate
Trends in bias and resolution across different shower geometries are consistent with CoREAS
The model has been successfully transferred via fine-tuning to the LOFAR experiment (different atmospheric and experimental conditions) using only a small amount of additional data

Highlights & Insights¶

Problem reformulation is the key contribution: recasting the stochastic microscopic simulation problem as a deterministic macroscopic regression task fundamentally reduces modeling complexity. This insight has broader implications for surrogate modeling in physical simulations.
Minimalist design: an 8-layer fully connected network with a 19 MB footprint, employing no elaborate architectural tricks, yet achieving physically meaningful high accuracy.
Differentiability dividend: the neural network surrogate is inherently differentiable, enabling integration with frameworks such as Information Field Theory (IFT).
Transfer learning viability: a model trained on one experiment can be fine-tuned to another with only a small amount of data.

Limitations & Future Work¶

Simulation accuracy is lower for weak pulses (low signal-to-noise regions), though the authors argue these regions are masked by noise in practice
Current training and validation are limited to the AERA frequency band (30–80 MHz) and associated atmospheric conditions
\(X_{\text{max}}\) reconstruction employs a simplified procedure (single scale factor \(S=1\)); the actual analysis pipeline is more complex
The longitudinal shower profile is not included as an additional input feature
Training data originate from a specific collaboration and are limited in volume
A bias exists at the low-\(X_{\text{max}}\) end of the sampling due to an asymmetric number of simulations on either side during parabolic fitting

Calorimeter simulation acceleration (CaloFlow, CaloDiffusion, CaloScore, etc.) employs generative models to accelerate particle physics detector simulations, but faces challenges in achieving high fidelity.
The hybrid approach adopted in this work—using conventional Monte Carlo for stochastic particle showering and a neural network for deterministic radio emission—represents a pragmatic and efficient strategy.
Inspiration from point cloud techniques: treating radio emission as a point-cloud prediction over antenna positions enables geometry-agnostic generalization.
The method has direct applicability to the radio astronomy community (AERA, LOFAR, SKA, GRAND, and related experiments).

Rating¶

Novelty: ⭐⭐⭐ — The methodology is relatively straightforward; the primary contribution lies in the problem reformulation insight and empirical validation.
Experimental Thoroughness: ⭐⭐⭐⭐ — Comprehensive evaluation covering pulse quality, fluence accuracy, and \(X_{\text{max}}\) reconstruction bias/resolution.
Writing Quality: ⭐⭐⭐⭐ — Concise and clear, with well-motivated physical reasoning.
Value: ⭐⭐⭐⭐ — Reducing simulation time from weeks to milliseconds represents substantial practical value for experimental physics.