Deep Ensemble Graph Neural Networks for Probabilistic Cosmic-Ray Direction and Energy Reconstruction in Autonomous Radio Arrays

Paper: 2602.23321 Authors: Arsène Ferrière, Aurélien Benoit-Lévy, Olivier Martineau-Huynh, Matías Tueros Categories: astro-ph.IM, cs.LG

Abstract

This paper presents a novel approach for reconstructing the arrival direction and energy of ultra-high-energy cosmic rays (UHECRs) using graph neural networks applied to radio detector arrays. The method represents triggered antennas as graph structures and leverages physics-informed architecture to achieve exceptional precision: 0.092° angular resolution and 16.4% electromagnetic energy reconstruction resolution under realistic noise conditions. The approach incorporates uncertainty estimation techniques to quantify prediction confidence and investigates robustness strategies for handling domain shifts between simulation and real-world observations.

Key Contributions

• Physics-informed graph neural network architecture that represents radio detector arrays as graph structures, reducing training data requirements while improving reconstruction precision
• State-of-the-art performance metrics: 0.092° angular resolution and 16.4% energy reconstruction resolution on simulated data with realistic noise
• Comprehensive uncertainty quantification framework providing confidence intervals for both direction and energy predictions
• Robustness analysis strategies to identify reliable prediction scenarios despite simulation-to-reality domain shifts
• Demonstration that incorporating physical knowledge into both architecture and input features significantly outperforms purely data-driven approaches

Methodology and Architecture

The core innovation lies in representing the radio detector array as a graph where each triggered antenna becomes a node. This graph structure naturally captures the spatial relationships and signal propagation patterns that are fundamental to cosmic-ray air shower physics.

The GNN architecture processes voltage traces from triggered antennas, with each node containing features derived from the measured signals. By encoding physical priors—such as the expected wavefront geometry and signal attenuation patterns—into both the network structure and input representations, the model requires substantially less training data than black-box approaches while achieving superior performance.

The ensemble approach combines multiple GNN models to generate probabilistic predictions. This ensemble strategy serves dual purposes: improving overall accuracy through model averaging and enabling robust uncertainty estimation through prediction variance analysis.

Uncertainty Quantification and Reliability

A critical aspect of this work is the emphasis on uncertainty estimation, which is essential for scientific applications where understanding prediction confidence is as important as the predictions themselves. The paper employs ensemble-based uncertainty quantification to provide:

• Confidence intervals for reconstructed arrival directions, allowing researchers to assess angular uncertainty on a per-event basis
• Energy reconstruction confidence bounds that reflect both statistical and systematic uncertainties
• Calibrated probability estimates that enable threshold-based event selection for downstream analyses

This probabilistic framework allows autonomous radio arrays to make informed decisions about which events warrant further investigation and which predictions can be trusted for scientific analysis.

Domain Shift and Robustness Analysis

One of the most practically important contributions is the investigation of model robustness under real-world variations. The authors recognize that simulations, however sophisticated, cannot perfectly capture all aspects of actual detector operation and environmental conditions.

The paper explores strategies to identify when the model’s predictions remain reliable despite differences between training simulations and deployment conditions. This includes analyzing prediction uncertainty patterns, examining feature distributions, and developing consistency checks that can flag potentially unreliable reconstructions.

This robustness analysis is crucial for autonomous operation, where the system must recognize its own limitations and potentially defer to alternative reconstruction methods or flag events for manual review when confidence is low.

Performance and Implications

The achieved angular resolution of 0.092° represents a significant advancement in cosmic-ray direction reconstruction from radio signals. This precision enables more accurate source localization studies and improved correlation analyses with other astronomical observations.

The 16.4% energy reconstruction resolution, while impressive, is particularly valuable when combined with the uncertainty estimates. Rather than treating all reconstructions equally, downstream analyses can weight events by their predicted reliability or apply confidence-based selection criteria.

For autonomous radio arrays deployed in remote locations, this approach offers the potential for real-time event classification and selective data transmission, reducing bandwidth requirements while maintaining scientific data quality.

Takeaways

1. Graph neural networks provide a natural and effective framework for processing spatially distributed detector arrays, with physics-informed architectures significantly outperforming purely data-driven approaches
2. Ensemble methods enable both improved accuracy and robust uncertainty quantification, essential for scientific applications requiring confidence estimates
3. The achieved 0.092° angular resolution and 16.4% energy resolution represent state-of-the-art performance for radio-based cosmic-ray reconstruction
4. Incorporating physical knowledge into machine learning models reduces training data requirements while improving generalization and interpretability
5. Robustness analysis and uncertainty quantification are critical for deploying ML models in autonomous scientific instruments where domain shift between simulation and reality is inevitable