Discrimination of Cosmic Rays and Neutrinos at Large Zenith Angles Using Machine Learning

Fengrong Zhu, Y. Z. Fan, M. M. Kang, Qi Gao2025 · DOI

TL;DR

This paper presents a machine learning approach to distinguish between cosmic ray showers and neutrino-induced showers at large zenith angles. By employing a stacking ensemble model trained on simulated data, the researchers achieve highly effective discrimination, paving the way for improved neutrino detection in future observatories.

What Problem Does This Paper Address?

Detecting high-energy astrophysical neutrinos is crucial for understanding the most energetic phenomena in the universe. However, neutrinos interact very weakly with matter, making their detection incredibly challenging. Cosmic rays, which are much more abundant and interact more readily, can mimic neutrino signals, creating a significant background. This is particularly true for neutrinos arriving at large zenith angles (close to the horizon), where their showers traverse a longer path through the Earth's atmosphere, potentially leading to more complex and harder-to-distinguish signals. Traditional methods for separating these signals often struggle with the high dimensionality and complexity of the data. This paper aims to address this challenge by developing a robust machine learning classifier to effectively discriminate between neutrino and cosmic ray (proton) induced air showers, especially in the challenging regime of large zenith angles.

Methods

The core of this work lies in a stacking-based machine learning approach. This ensemble method combines multiple "base learners" to make predictions, and then uses a "meta-learner" to combine the outputs of these base learners.

Here's a breakdown:

  • Base Learners: The researchers trained three different classifiers:
  • Random Forest: Known for its ability to reduce overfitting by combining multiple decision trees.
  • XGBoost: A powerful gradient boosting algorithm that iteratively optimizes its predictions.
  • Support Vector Machine (SVM): Utilizes the RBF kernel to model complex, non-linear decision boundaries.
  • Meta-Learner: The outputs from these three base learners are then fed into a Logistic Regression classifier, which acts as the meta-learner to make the final prediction. This architecture aims to leverage the complementary strengths of each base model for improved generalization.
  • Feature Engineering: The model is trained on a set of 13 features extracted from simulated air showers. These features characterize both the electromagnetic and muonic components of the showers, including statistical moments of their lateral distributions, NKG fit parameters, particle counts, and the electron-to-muon ratio.

Datasets

The study utilizes simulated data generated using the CORSIKA software. This is a standard tool for simulating extensive air showers.

  • Particle Types: Both neutrinos and protons (as a proxy for cosmic rays) were simulated.
  • Energy Ranges: Three distinct energy intervals were considered: 100 TeV–1 PeV, 1 PeV–10 PeV, and 10 PeV–100 PeV.
  • Zenith Angles: Four large zenith angles were investigated: 60°, 70°, 80°, and 89°.
  • Event Generation: For each combination of energy and zenith angle, 5000 events were generated.

The paper does not introduce a new benchmark dataset but rather uses simulated data to test and validate their proposed classification method.

Key Insights and Ideas

  • Machine Learning for Neutrino Astrophysics: The study strongly validates the utility of machine learning, specifically ensemble methods like stacking, for tackling the challenging problem of neutrino detection against a dominant cosmic ray background.
  • Feature Importance at Large Zenith Angles: The research suggests that the features extracted from air showers become particularly powerful discriminators at very large zenith angles (e.g., 89°). This is a crucial finding, as these are the angles where neutrino signals are often harder to isolate.
  • Stacking Ensemble Superiority: The stacking approach, by combining multiple diverse models, demonstrates superior performance compared to individual classifiers, highlighting the benefit of ensemble learning in complex classification tasks.
  • Effective Signal-Background Separation: The high AUC and Q-factor values achieved indicate that the model is highly effective at separating true neutrino signals from background proton events, even in challenging scenarios.

Key Results

The stacking-based machine learning model demonstrated strong performance across all tested energy and zenith angle bins:

  • Classification Performance:
    • AUC (Area Under the ROC Curve): Generally exceeded 0.94, with many bins reaching above 0.99. An AUC close to 1 signifies excellent classification.
  • Q-factor: Consistently above 96, and in some cases reaching over 9000 (e.g., 4097.15 for 1 PeV-10 PeV at 70°). A higher Q-factor indicates better signal-background separation.
  • Generalization: The difference between training and validation AUC values was typically less than 0.03, indicating good generalization ability and minimal overfitting.
  • Detection Significance: The estimated detection significance (Z) exceeded 4.5σ in most energy-angle bins under realistic observational conditions, suggesting a high confidence level for neutrino detection with this method.

Limitations and Future Work

  • Limitations:
    • The study relies entirely on simulated data. While CORSIKA is a well-established tool, real-world detector effects and uncertainties are not fully captured.
  • The analysis focuses on specific energy ranges and zenith angles. Performance at other energies or angles might differ.
  • Future Work:
    • The authors suggest that this work provides a reference framework for future neutrino observations.
  • Applying this method to real data from existing or future neutrino observatories (like IceCube, KM3NeT, or Baikal-GVD) is the natural next step to validate its performance in practice.
  • Further exploration of additional features or more advanced machine learning architectures could potentially improve performance even further.

Summary updated

Abstract

Simulated data of neutrinos and protons were generated using the CORSIKA software across three energy ranges—100 TeV–1 PeV, 1 PeV–10 PeV, and 10 PeV–100 PeV—and four zenith angles: 60◦ , 70◦ , 80◦ , and 89◦ . For each event, secondary particles within 0–600 m from the shower axis were analyzed. Electromagnetic and muonic components were characterized using statistical moments of lateral distributions, NKG fit age parameters, particle counts, and the electron-to muon ratio. A stacking model was employed to build a classifier to distinguish between neutrino and proton events. Features were input into the model separately for each energy-angle bin. The classification performance was evaluated using the AUC score and 𝑄-factor. Results show strong classification capability across various energy and angle combinations, with AUC values generally exceeding 0.94 and 𝑄-factors reaching above 96, indicating effective signal-background separation. Further significance analysis incorporating flux estimates revealed that the detection significance exceeds 4.5𝜎 in most energy-angle bins, demonstrating the method’s potential for neutrino–proton discrimination and providing a reference framework for future neutrino observations aimed at exploring their astrophysical origins.
BibTeX
@misc{zhu2025discrimination,
  author = {Fengrong Zhu and Y. Z. Fan and M. M. Kang and Qi Gao},
  title = {{Discrimination of Cosmic Rays and Neutrinos at Large Zenith Angles Using Machine Learning}},
  year = {2025},
  doi = {10.22323/1.501.1230},
}
Figure