Simulation-Based Inference for Direction Reconstruction of Ultra-High-Energy Cosmic Rays with Radio Arrays

This paper presents a simulation-based inference pipeline combining a physics-informed graph neural network with a normalizing flow to achieve sub-degree angular resolution and well-calibrated uncertainties for reconstructing the arrival directions of ultra-high-energy cosmic rays using sparse radio arrays.

The Gist

Imagine the universe is throwing a massive, invisible party. The guests are Ultra-High-Energy Cosmic Rays (UHECRs)—particles so energetic they could power a city for a day, yet they are so rare and elusive that we barely know where they come from.

To find these guests, scientists use giant "radio antennas" scattered across deserts (like the planned GRAND array). When a cosmic ray hits the atmosphere, it creates a massive cascade of particles (an "air shower") that hits the ground. As this shower moves, it emits a tiny, nanosecond-long radio pulse. The antennas catch this pulse, and the goal is to figure out exactly where in the sky the original particle came from.

The Problem: The "Guessing Game" Trap

Traditionally, scientists tried to reconstruct the direction by fitting a simple mathematical shape (like a flat sheet of paper) to the timing of the radio pulses hitting the antennas.

• The Analogy: Imagine trying to guess the direction of a marching band by listening to when the drumbeat hits your ears at different spots in a park. If you assume the sound travels in a perfectly flat wave, you might get close. But in reality, the sound waves curve, the wind blows, and the band is moving.
• The Issue: These old "flat sheet" models are too simple. They often give a "best guess" direction but fail to tell you how confident they are. They might say, "It's 99% sure it's here," when in reality, the uncertainty is huge. This is dangerous for multi-messenger astronomy, where you need to point other telescopes at the right spot to catch a neutrino or a gamma ray.

The Solution: The "Smart Detective" Pipeline

This paper introduces a new, super-smart method called Simulation-Based Inference (SBI). Instead of trying to write a perfect math formula for the universe, they taught a computer to "learn" from millions of simulated parties.

Here is how their new pipeline works, step-by-step:

1. The "Rough Sketch" (The Physics Seed)

First, the system takes a quick, simple guess using the old "flat sheet" math.

• Analogy: This is like a detective looking at a crime scene and saying, "Okay, based on the footprints, the suspect probably came from the north." It's a good starting point, but it's not perfect.

2. The "Pattern Recognizer" (The Graph Neural Network)

Next, a specialized AI called a Graph Neural Network (GNN) looks at the data. It treats every antenna as a "node" in a web and looks at the connections between them.

• Analogy: Imagine the detective now calling in a team of experts who know the local terrain. They look at the relationships between the footprints: "Wait, the mud here is deeper, and the time gap between these two spots is weird. The suspect didn't walk in a straight line; they zig-zagged."
• The GNN learns the complex, messy patterns of the radio waves that the simple math missed. It doesn't just guess a direction; it learns the shape of the wavefront.

3. The "Uncertainty Calculator" (The Normalizing Flow)

This is the magic part. Instead of just giving one answer, the system outputs a full map of possibilities (a Bayesian posterior).

• Analogy: Instead of the detective saying, "The suspect is at the North Gate," they draw a circle on a map.
  • The Old Way: The circle is tiny, but it often misses the suspect (over-confident).
  • The New Way: The AI draws a circle that is slightly larger, but it is statistically guaranteed to contain the suspect 68% of the time. It knows exactly how "fuzzy" the answer is.

4. The "Reality Check" (Temperature Calibration)

The scientists noticed their AI was being a little too cautious (drawing circles that were slightly too big). So, they applied a "temperature" adjustment.

• Analogy: Think of the AI's confidence like a thermostat. If it's too cold (too broad), they turn up the heat slightly to tighten the circle until it hits the perfect "Goldilocks" zone: big enough to be accurate, but small enough to be useful.

Why This Matters

• Speed: It works incredibly fast, processing data in milliseconds.
• Trust: It provides a "confidence score" that is mathematically proven to be correct. If it says "68% chance," you can bet your life on it.
• Future-Proof: This method is ready for the next generation of massive radio arrays (like GRAND, AugerPrime, and BEACON) that will cover thousands of square kilometers.

The Bottom Line

The authors built a hybrid detective: one part uses the laws of physics for a quick start, and the other part uses deep learning to understand the messy details. The result is a system that doesn't just point a finger at the sky; it draws a target that is guaranteed to hit the bullseye, helping us finally solve the mystery of where these ultra-powerful cosmic particles come from.

Technical Summary

Here is a detailed technical summary of the paper "Simulation-Based Inference for Direction Reconstruction of Ultra-High-Energy Cosmic Rays with Radio Arrays."

1. Problem Statement

Ultra-high-energy cosmic rays (UHECRs) are rare, energetic particles whose origins remain unknown. Accurate reconstruction of their arrival directions is critical for multi-messenger astronomy and composition studies.

• The Challenge: Current radio detection arrays (e.g., GRAND, AugerPrime, BEACON) rely on fitting geometric wavefront models (e.g., plane or spherical) to antenna trigger times. While effective, these explicit likelihood methods often rely on simplified physical assumptions (e.g., planar wavefronts) that can introduce biases and fail to rigorously quantify uncertainties.
• The Limitation: Traditional Bayesian inference using Markov Chain Monte Carlo (MCMC) is computationally prohibitive for large arrays with complex wavefront structures and non-Gaussian noise. Furthermore, standard algorithms often provide only a "best-fit" direction without a full, calibrated posterior distribution, making uncertainty quantification difficult for downstream science.

2. Methodology

The authors propose a Simulation-Based Inference (SBI) pipeline, also known as likelihood-free inference, implemented within the Learning the Universe Implicit Likelihood Inference (LtU-ILI) framework. The approach bypasses the need for an analytic likelihood function by learning the posterior distribution directly from high-fidelity simulations.

A. Data Generation

• Simulator: The pipeline uses ZHAireS, a physics-based code that simulates extensive air showers (EAS) and calculates radio emission based on first principles (Liénard-Wiechert potentials).
• Dataset: ~8,200 realistic UHECR events (protons and iron nuclei) with energies from 0.4 to 4 EeV and zenith angles $37^\circ \leq \theta \leq 87^\circ$.
• Observables: The input data consists of antenna coordinates $(x, y, z)$ and GPS-synchronized trigger times $(t)$. The trigger times are derived from the Hilbert envelope of the electric field traces, with added Gaussian jitter ($\sigma = 5$ ns) to mimic GPS synchronization uncertainty.

B. Neural Architecture

The pipeline employs a hybrid architecture combining physics-informed analytical models with deep learning:

1. Graph Neural Network (GNN) Encoder:
   • Each air shower event is modeled as a graph where nodes are triggered antennas and edges connect $k$-nearest neighbors in space-time.
   • The GNN learns spatiotemporal correlations among antenna signals to predict a residual correction vector ($\Delta k$) to the arrival direction.
2. Physics-Informed Prior (Plane Wavefront - PWF):
   • An analytic PWF fit provides an initial direction estimate ($k_{PWF}$). This ensures the solution respects causal ordering and global geometry.
3. Residual Gating Mechanism:
   • A learned scalar weight ($\alpha \in [0, 1]$) blends the PWF estimate and the GNN correction: $k_{final} = k_{PWF} + \alpha \Delta k$. This allows the network to apply non-linear corrections only where necessary, maintaining physical realism.
4. Normalizing Flow (Posterior Estimator):
   • The output of the encoder conditions an 8-block Masked Autoregressive Flow (MAF).
   • The MAF transforms a simple Gaussian base distribution into the complex, full Bayesian posterior $p(k | \text{data})$.
   • This allows for rapid sampling of the full posterior distribution (directional uncertainty) without iterative optimization.

C. Calibration

• Temperature Scaling: To ensure the posterior is statistically well-calibrated (i.e., a 68% credible region actually contains the true direction 68% of the time), the authors apply a post-hoc temperature scaling ($T$) to the log-density: $p_T \propto p^{1/T}$.
• Optimization: $T$ is tuned on a held-out calibration set to match empirical coverage targets.

3. Key Contributions

1. First End-to-End SBI Pipeline for UHECR Radio Arrays: This work demonstrates a fully simulation-based approach for direction reconstruction that avoids explicit likelihood modeling.
2. Hybrid Physics-ML Architecture: The integration of a GNN with a physics-informed PWF prior via a residual gating mechanism balances the flexibility of deep learning with the interpretability and robustness of analytical models.
3. Rigorous Uncertainty Quantification: Unlike traditional methods that provide point estimates, this pipeline outputs a full, calibrated Bayesian posterior, enabling robust multi-messenger follow-up.
4. Validation Framework: The use of TARP (Tests of Accuracy with Random Points) and P-P plots provides a rigorous statistical validation of the posterior calibration, proving the method is not over-confident.

4. Results

The pipeline was tested on an independent validation set of 1,560 simulated GRANDproto300 events.

• Angular Resolution:
  • Achieved a median 68% containment angle ($R_{68\%}$) of $0.38^\circ$.
  • Only ~3% of events had an error exceeding $1^\circ$.
  • Resolution improves for highly inclined showers ($\theta > 70^\circ$), reaching a median of $\approx 0.30^\circ$.
• Calibration Performance:
  • The nominal 68% highest-posterior-density (HPD) contours captured $71% \pm 2%$ of the true arrival directions.
  • This indicates a mildly conservative uncertainty calibration, which is desirable in astrophysics to avoid underestimating errors.
  • TARP diagnostics confirmed the posteriors are well-calibrated across the entire credibility range.
• Dependence on Event Geometry:
  • Zenith Angle: Resolution improves as the zenith angle increases (more inclined showers).
  • Antenna Multiplicity: Resolution improves monotonically with the number of triggered antennas ($N_{trig}$). For $N_{trig} \geq 30$, the error plateaus near $0.30^\circ$, limited by the 5 ns timing jitter rather than geometry.
• Comparison with Analytic Methods:
  • While the analytic PWF method can achieve slightly better point-estimate resolution ($\approx 0.2^\circ$) in some regimes, it produces over-confident uncertainties (empirical coverage as low as 2–12% for a nominal 68% interval).
  • The SBI method trades a negligible increase in point-estimate error for statistically valid, calibrated uncertainty regions.

5. Significance and Future Outlook

• Multi-Messenger Astronomy: The ability to provide rapid, calibrated, sub-degree direction estimates is crucial for identifying sources of UHECRs and neutrinos, particularly for upcoming large-scale radio arrays like GRAND, AugerPrime Radio, and BEACON.
• Scalability: The amortized nature of the SBI approach (inference time is constant regardless of event complexity) makes it ideal for real-time processing of massive datasets from next-generation observatories.
• Future Work: The authors suggest extending the input features to include signal amplitudes and polarization to simultaneously reconstruct energy and shower depth ($X_{max}$), and integrating these methods into hierarchical flows for joint composition analysis.

In conclusion, this paper establishes a new standard for UHECR reconstruction by demonstrating that simulation-based inference can deliver physically interpretable, sub-degree accurate, and rigorously calibrated direction estimates, overcoming the limitations of traditional likelihood-based approaches.