🔬 physics

Search for Cosmic Rays in GRANDProto300

This paper presents the cosmic-ray search pipeline and a developing CNN-based identification algorithm for the GRANDProto300 prototype array, detailing their filtering steps, efficiency evaluation on data from July to October 2025, and model accuracy in distinguishing cosmic-ray signals from noise.

Original authors: Jolan Lavoisier, Aurélien Benoit-Lévy, Kumiko Kotera, Mauricio Bustamante, Marion Guelfand, Nathan Lebas, Takashi Sako, Xishui Tian, for the GRAND Collaboration

Published 2026-10-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Jolan Lavoisier, Aurélien Benoit-Lévy, Kumiko Kotera, Mauricio Bustamante, Marion Guelfand, Nathan Lebas, Takashi Sako, Xishui Tian, for the GRAND Collaboration

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

High above the Earth, the atmosphere is constantly being bombarded by invisible particles traveling at nearly the speed of light. These are cosmic rays, energetic messengers from deep space that strike our planet's air, creating showers of secondary particles that ripple down to the ground. While scientists have long studied these particles, there is a specific energy range where their origins remain a mystery. It is the boundary where particles born within our own Milky Way galaxy might give way to those arriving from distant, powerful sources across the universe. Pinpointing this transition is crucial for understanding how the cosmos accelerates matter to such extreme energies, but detecting these specific particles is difficult because they arrive rarely and leave only fleeting, faint traces.

To solve this puzzle, a team of researchers is building a massive listening post in the Gobi Desert of China. This project, known as GRAND, uses a vast array of radio antennas to catch the faint radio signals produced when cosmic rays hit the atmosphere. The paper describes the progress of GRANDProto300, a prototype version of this experiment designed to prove that this radio-detection method works. Between late 2024 and mid-2025, the team operated a smaller version of the array, eventually deploying 65 antennas across the desert, with 63 of them operational and collecting data stably. Their goal was not just to listen, but to teach their computers how to distinguish the genuine signal of a cosmic ray from the overwhelming noise of the environment, such as radio waves from aircraft or electrical equipment.

The researchers faced a significant challenge: the raw data collected by the antennas was flooded with false alarms. To find the real cosmic rays, they developed a step-by-step filtering process, much like sifting through a large pile of sand to find a few specific grains. First, they removed bursts of noise that appeared in clusters, which often came from commercial airplanes or a nearby electrical transformer. Next, they looked at the direction of the radio waves. Because cosmic rays interact with Earth's magnetic field, they produce signals with a specific orientation that differs from random background noise. They also filtered out signals coming from straight overhead or nearly flat along the horizon, as the geometry of their antenna array made it difficult to accurately reconstruct those specific angles. Finally, they checked the timing of the signals across different antennas to ensure they matched the expected travel time of a wave coming from a single point in the sky.

After applying these strict filters to millions of recorded events, the team narrowed the list down to just 51 strong candidates that looked like genuine cosmic rays. These candidates showed the expected patterns of arrival and signal strength, suggesting the filtering method was working. However, the team knew that manual rules might not catch everything, so they began training a computer program using a type of artificial intelligence called a neural network. This program learned to recognize the unique shape of a cosmic ray signal by studying thousands of examples from both computer simulations and real data. When tested, this new tool proved highly effective at identifying simulated signals, correctly flagging nearly all of them. Its performance on real, recorded cosmic ray signals was good but slightly lower, indicating that while the method is powerful, there is still room to refine how it handles the complexities of real-world data.

The work presented here represents a vital step forward in the search for the origins of high-energy cosmic rays. By successfully operating the prototype array and developing a pipeline that can separate true cosmic events from noise, the researchers have validated the core technique for the larger GRAND experiment. The identification of 51 candidates and the development of a machine learning tool to improve future detection show that the path is clear. The team is now moving toward a larger, more sensitive array that will be able to map the transition between galactic and extragalactic cosmic rays, potentially revealing the hidden engines of the universe that accelerate these particles to such incredible speeds.

Technical Summary: Cosmic Ray Search in GRANDProto300

Problem Statement
The GRANDProto300 (GP300) experiment, a prototype array of the Giant Radio Array for Neutrino Detection (GRAND), aims to validate the autonomous radio-detection technique for astroparticles. Specifically, it targets cosmic rays with energies between $10^{17}$ and $10^{18.5}$ eV to study the transition region between Galactic and extragalactic sources. The primary challenge addressed in this work is the identification of genuine cosmic-ray signals within a massive dataset dominated by noise. The data is subject to two types of interference: "parasites" (transient noise causing false triggers, such as from aircraft or electric transformers) and stationary background noise. The goal is to develop a robust pipeline to filter these events and distinguish true cosmic-ray air showers from noise with high purity.

Methodology
The authors present a two-pronged approach: a deterministic filtering pipeline and a machine learning-based identification layer.

Deterministic Search Pipeline:
The pipeline applies a series of conservative cuts to a dataset of 5,477,117 Coincidence Data (CD) events recorded between July and October 2025. The process prioritizes purity over efficiency:
- Cluster Cut: Removes events clustered in time and space (angular distance $\Delta\vartheta < 5^\circ$ , temporal distance $\Delta t < 5$ s) to eliminate interference bursts. This step removes 88% of the data.
- Polarization Cut: Utilizes the geomagnetic field vector ( $B$ ) and the electric field vector ( $e_V$ ). Since cosmic-ray signals are expected to be orthogonal to $B$ ( $e_V \cdot e_B \approx 0$ ), events failing this criterion are rejected, removing 35% of events.
- Zenith Cut: Removes events with reconstructed zenith angles $\theta \leq 60^\circ$ (low efficiency for the layout) and $\theta \geq 88^\circ$ (reconstruction difficulties), eliminating 61% of the batch.
- Delay Cut: Uses Spherical Wave Fit (SWF) reconstruction to compare measured arrival times against theoretical expectations. Events with a high $\chi^2$ discrepancy (threshold $10^3$ ) are filtered, removing 44% of the data.
- Quality Cuts: A final Angular Distribution Fit (ADF) with a $\chi^2_{ADF} < 20$ threshold is applied.
- Outcome: These cuts collectively filter out 99.63% of parasites, leaving approximately 20,000 candidates for further analysis, from which 51 cosmic-ray candidates were selected.
Machine Learning Identification Layer:
To address the shape of recorded traces (which the deterministic cuts do not fully utilize), the authors developed a Convolutional Neural Network (CNN) for binary classification (noise vs. cosmic-ray).
- Dataset Construction: Traces were extracted from July 2025 data. The dataset includes 384-point traces from real parasites, simulated cosmic rays (inserted into background noise), and 567 real cosmic-ray traces identified via the pipeline.
- Architecture: The model uses 13 convolutional layers with batch normalization and ReLU activation, followed by global average pooling and a dense layer. Two variants were tested: 1D and 2D convolutions, processing X (North-South) and Y (East-West) channels.
- Training: Models were trained for 4000 epochs with a batch size of 128 and a learning rate of 0.01, optimizing for recall (True Positives / (True Positives + False Negatives)).

Key Contributions

Operational Validation: The paper reports on the stable operation of GP300, with 63 antennas collecting data as of July 2025, and presents the first systematic cosmic-ray search pipeline for this prototype.
Filtering Efficiency: The authors demonstrate a pipeline capable of reducing a dataset of over 5 million events to a manageable set of candidates by removing 99.63% of noise events through systematic physical and geometric cuts.
ML Integration: The paper introduces a CNN-based algorithm specifically designed to discriminate between noise and cosmic-ray signals using signal traces, marking a step toward more sophisticated offline analysis.
Candidate Identification: The pipeline successfully identified 51 cosmic-ray candidates, whose arrival directions and signal dominance align with theoretical expectations regarding the geomagnetic field.

Results

Pipeline Performance: The deterministic cuts reduced the initial 5.47 million events to 51 candidates. The arrival directions of these candidates (visualized in Figure 1) show the expected East-West signal dominance.
Machine Learning Performance:
- Simulated Traces: The models achieved high recall rates of 95.71% (1D model) and 97.85% (2D model).
- Real Cosmic-Ray Traces: The recall was lower but significant, at 62.08% (1D) and 71.15% (2D).
- Parasite Rejection: The models effectively distinguished parasites, with the 2D model showing a recall of 71.15% for real cosmic rays while maintaining a low false positive rate on the test set of parasites.

Significance and Claims
The paper claims that the developed pipeline successfully validates the technique of autonomous radio-detection for cosmic rays in the $10^{17} - 10^{18.5}$ eV range. By combining systematic physical cuts with a developing machine learning layer, the authors demonstrate the feasibility of isolating cosmic-ray signals from a high-noise environment. The identification of 51 candidates serves as a proof-of-concept for the GRAND experiment's future large-scale operations. The authors modestly note that while visual selection criteria were temporarily applied, the systematic development of trace-shape cuts (via the CNN) is an ongoing effort to improve identification without relying on manual inspection. The work establishes a foundation for the study of the Galactic-to-extragalactic transition region using the GRAND array.

Technical Summary: Cosmic Ray Search in GRANDProto300

More like this