Radio signal generation in milliseconds: enabling multi-parameter reconstruction of ultra-high-energy cosmic rays

This paper introduces a machine-learning-based emulator that generates ultra-high-energy cosmic ray radio signals in milliseconds with high accuracy, enabling efficient multi-parameter reconstruction of primary particle properties that matches state-of-the-art performance on both simulated and real data from the GRANDProto300 experiment.

The Gist

Imagine the universe is constantly raining down invisible, super-fast particles called Ultra-High-Energy Cosmic Rays (UHECRs). When these particles hit Earth's atmosphere, they don't just stop; they crash into air molecules and create a massive, expanding explosion of secondary particles, known as an "air shower."

As this shower spreads out, the charged particles inside it wiggle through Earth's magnetic field. This wiggling creates a faint radio signal, like a tiny lightning bolt that you can't see but can "hear" with the right equipment.

The paper describes a new, super-fast way to decode these radio signals to figure out exactly what kind of cosmic ray started the party and where it came from. Here is the breakdown of their invention:

1. The Problem: The "Slow Cooker" vs. The "Microwave"

Traditionally, scientists use complex computer programs (called ZHAireS and CoREAS) to simulate how these cosmic ray showers behave and what their radio signals should look like.

• The Old Way: Think of these simulations like a slow cooker. To get one accurate result, the computer has to "stir" the simulation for several hours. If you want to figure out the properties of a cosmic ray by comparing real data to millions of possible simulations (a method called Bayesian reconstruction), you would need to run the slow cooker millions of times. That would take years!
• The New Way: The authors built a Machine Learning Emulator. Think of this as a "microwave" or a "smart shortcut." It has studied millions of those slow-cooker simulations and learned the patterns. Now, instead of taking hours, it can predict what the radio signal should look like in just milliseconds (a thousandth of a second).

2. How the "Smart Shortcut" Works

The machine learning model is like a very talented translator.

• The Input: You give it the "recipe" of the cosmic ray: Where did it come from? How much energy did it have? How deep did it go into the atmosphere?
• The Output: It instantly tells you what the radio signal looks like.
• The Trick: Instead of trying to memorize every single wiggle of the radio wave (which is like trying to memorize every pixel in a photo), the model learns to describe the wave using just five simple numbers (like the height, width, and shape of a hill). This makes the math much faster and easier.

3. The Result: A Crystal Clear Picture

The team tested this "microwave" against the "slow cooker" (the real simulations).

• Accuracy: The emulator was incredibly accurate. The difference between its prediction and the real simulation was only about 5%. This is good enough that it's actually better than the difference between the two different slow-cooker programs scientists usually use!
• Reconstruction: They used this fast emulator to look at real data from the GP300 prototype (a radio telescope array in China). By comparing the real radio signals to the emulator's predictions, they could figure out:
  • Energy: How powerful the cosmic ray was (within 8.9% accuracy).
  • Direction: Where it came from in the sky (within 0.08 degrees accuracy—imagine hitting a bullseye from a mile away).

4. The Real-World Test

Finally, they didn't just test it on fake data. They took 32 real cosmic ray candidates detected by the GP300 prototype and ran them through their new system.

• The results matched perfectly with the older, slower methods used by the same team.
• This proves that the "microwave" works just as well as the "slow cooker" but is fast enough to be useful for real-time science.

Summary

In short, the authors built a super-fast AI assistant that learned to predict cosmic ray radio signals. It turns a process that used to take hours into one that takes milliseconds, allowing scientists to reconstruct the history of these cosmic particles with high precision, all while using real data from a prototype telescope.

Technical Summary

1. Problem Statement

Ultra-high-energy cosmic rays (UHECRs), with energies exceeding $10^{18}$ eV, are detected via the radio emissions produced by extensive air showers (EAS) interacting with Earth's magnetic field. While Monte Carlo simulation codes like ZHAireS and CoREAS can model these radio signals with high accuracy, they are computationally expensive, typically requiring several hours to simulate a single event.

This high computational cost creates a bottleneck for Bayesian reconstruction of cosmic ray properties (energy, direction, and shower maximum depth). Bayesian approaches require sampling the input parameter space extensively to determine posterior distributions, a process that is infeasible with traditional hour-long simulations. There is a critical need for a method that can generate radio signal predictions in milliseconds without sacrificing accuracy.

2. Methodology

The authors propose a Machine Learning (ML) based emulator trained on ZHAireS simulations to predict radio signals instantly. The methodology consists of three main components:

A. Data Parametrization

To optimize the input and output spaces for the neural network, the authors introduced specific parametrizations:

• Input Parameters: Instead of absolute coordinates, the model uses relative positioning of antennas with respect to the shower maximum ($X_{max}$). The coordinate system is defined by the shower axis ($\vec{k}$) and the geomagnetic field ($\vec{B}$).
  • Inputs include: Zenith angle ($\theta$), Azimuth angle ($\phi$), off-axis angle ($\omega_i$), polar angle ($\eta_i$), antenna distance to $X_{max}$ ($l_i$), electromagnetic energy ($E$), and atmospheric depth parameters ($z$-component of $X_{max}$ and effective refractive index $n_{eff}$).
• Output Parametrization: To reduce dimensionality from 1024 time bins $\times$ 3 polarization channels to a manageable size, the radio signal is modeled in the frequency domain.
  • The signal is assumed to be dominated by geomagnetic emission (polarized along $\vec{k} \times \vec{B}$).
  • The Fourier transform of the signal, $S(f)$, is parametrized using 5 parameters:
    • Amplitude: $|S(f)| = \exp(a + b(f-f_0) + c(f-f_0)^2)$
    • Phase: $\Phi(f) = \Phi_0 + \Phi_p(f-f_0) + \Phi_q(f-f_0)^2$
  • This reduces the output space to just 5 parameters per antenna, valid up to a frequency $f_{thin}$ (where simulation noise from particle thinning becomes dominant).

B. Model Architecture and Training

• Architecture: A feedforward neural network with 4 hidden layers, each containing 300 neurons.
• Training Data: 15,000 ZHAireS simulations of proton and iron showers at the GRANDProto300 (GP300) site in Gansu, China.
• Loss Function: Mean Squared Error (MSE) between predicted and true parameters.
• Inference Speed: The model predicts the electric field signal in milliseconds. Converting this to voltage traces (using antenna response modeling) takes an average of 6 milliseconds for an event with 20 antennas on a standard personal computer.

C. Bayesian Parameter Reconstruction

The emulator is integrated into a Markov Chain Monte Carlo (MCMC) framework to reconstruct cosmic ray properties:

• Likelihood Function: Defined as the product of the likelihood of voltage traces and signal timing.
  • Voltage likelihood compares measured traces to emulated traces, accounting for emulator error ($\sigma_s$) and measurement noise ($\sigma_n$).
  • Timing likelihood compares arrival times based on the speed of light and effective refractive index.
• Priors: Constraints are applied based on training data distributions, specifically the correlation between zenith angle and $X_{max}$ depth, and zero probability for parameter combinations outside the training set.

3. Key Contributions

1. Speed vs. Accuracy Trade-off: The development of an emulator that reduces signal generation time from hours to milliseconds while maintaining accuracy comparable to the discrepancy between different physics simulation codes (ZHAireS vs. CoREAS).
2. Efficient Parametrization: A novel frequency-domain parametrization of the radio signal that drastically reduces output dimensionality without losing physical information.
3. End-to-End Reconstruction Pipeline: A complete workflow combining ML-based signal generation with MCMC-based Bayesian inference, successfully applied to real-world prototype data.
4. Validation on Real Data: The first successful application of this ML-driven reconstruction method to actual cosmic-ray candidates detected by the GP300 prototype.

4. Results

The paper reports performance metrics based on simulations and real data:

• Signal Generation Accuracy:
  • Relative error on voltage amplitude: 5.2%
  • Relative error on fluence: 8.2%
  • Note: These errors are smaller than the intrinsic differences observed between ZHAireS and CoREAS simulations.
• Reconstruction Performance (Simulated Data):
  • Electromagnetic Energy Resolution: 8.9%
  • Angular Resolution: 0.08°
  • Atmospheric Depth ($X_{max}$) Resolution: 81 g/cm²
  • Alternative Method: Using only maximum amplitude (similar to Angular Distribution Function methods) yielded 12.5% energy resolution and 0.05° angular resolution.
• Real Data Application (GP300):
  • Successfully reconstructed 32 out of 51 cosmic-ray candidates.
  • The reconstructed energy spectrum and arrival directions were consistent with those obtained using the traditional Angular Distribution Function (ADF) method.

5. Significance

This work represents a paradigm shift in UHECR detection analysis. By replacing computationally intensive physics simulations with a fast, accurate machine learning emulator, the authors enable real-time or near-real-time Bayesian reconstruction. This capability is crucial for future large-scale radio arrays (like the full GRAND experiment) where the volume of data will be massive, and traditional simulation-based reconstruction would be prohibitively slow. The successful validation on GP300 prototype data confirms the method's viability for upcoming cosmic ray observatories, paving the way for high-precision multi-parameter reconstruction of ultra-high-energy events.