Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Using Deep Learning for Robust Classification of Fast Radio Bursts

Rohan Arni, Carlos Blanco, Anirudh Prabhu

TL;DR

This work addresses the challenge of classifying FRBs as repeaters or non-repeaters while uncovering the latent structure of FRB signals learned from the CHIME catalog. The authors introduce a supervised variational autoencoder that jointly optimizes reconstruction, latent-space regularization, and classification, yielding a compact, interpretable latent embedding. The model achieves high accuracy (F2 ≈ 0.9807) and reveals that dispersion measure excess, spectral index, and spectral running are key discriminators, with several repeater candidates identified among false positives. Limitations include instrumental biases from CHIME data and reliance on derived features; the approach remains promising for cross-catalog expansion and deeper physical interpretation.

Abstract

While the nature of fast radio bursts (FRBs) remains unknown, population-level analyses can elucidate underlying structure in these signals. In this study, we employ deep learning methods to both classify FRBs and analyze structural patterns in the latent space learned from the first CHIME catalog. We adopt a Supervised Variational Autoencoder (sVAE) architecture which combines the representational learning capabilities of Variational Autoencoders (VAEs) with a supervised classification task, thereby improving both classification performance and the interpretability of the latent space. We construct a learned latent space in which we perform further dimensionality reduction to find underlying structure in the data. Our results demonstrate that the sVAE model achieves high classification accuracy for FRB repeaters and reveals separation between repeater and non-repeater populations. Upon further analysis of the latent space, we observe that dispersion measure excess, spectral index, and spectral running are the dominant features distinguishing repeaters from non-repeaters. We also identify four non-repeating FRBs as repeater candidates, two of which have been independently flagged in previous studies.

Using Deep Learning for Robust Classification of Fast Radio Bursts

TL;DR

This work addresses the challenge of classifying FRBs as repeaters or non-repeaters while uncovering the latent structure of FRB signals learned from the CHIME catalog. The authors introduce a supervised variational autoencoder that jointly optimizes reconstruction, latent-space regularization, and classification, yielding a compact, interpretable latent embedding. The model achieves high accuracy (F2 ≈ 0.9807) and reveals that dispersion measure excess, spectral index, and spectral running are key discriminators, with several repeater candidates identified among false positives. Limitations include instrumental biases from CHIME data and reliance on derived features; the approach remains promising for cross-catalog expansion and deeper physical interpretation.

Abstract

While the nature of fast radio bursts (FRBs) remains unknown, population-level analyses can elucidate underlying structure in these signals. In this study, we employ deep learning methods to both classify FRBs and analyze structural patterns in the latent space learned from the first CHIME catalog. We adopt a Supervised Variational Autoencoder (sVAE) architecture which combines the representational learning capabilities of Variational Autoencoders (VAEs) with a supervised classification task, thereby improving both classification performance and the interpretability of the latent space. We construct a learned latent space in which we perform further dimensionality reduction to find underlying structure in the data. Our results demonstrate that the sVAE model achieves high classification accuracy for FRB repeaters and reveals separation between repeater and non-repeater populations. Upon further analysis of the latent space, we observe that dispersion measure excess, spectral index, and spectral running are the dominant features distinguishing repeaters from non-repeaters. We also identify four non-repeating FRBs as repeater candidates, two of which have been independently flagged in previous studies.

Paper Structure

This paper contains 20 sections, 12 equations, 4 figures, 9 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Data Preprocessing
  4. Model Architecture
  5. Training Procedure
  6. Evaluation Metrics
  7. Evaluation Metrics
  8. Latent Space Analysis
  9. Results
  10. Optimized Hyperparameters
  11. Optimized Model Layers
  12. Classifier Performance
  13. False Positive Analysis
  14. Latent Space Visualization
  15. Correlation Tests
  16. ...and 5 more sections

Figures (4)

  • Figure 1: Schematic of model architecture.
  • Figure 2: PCA and t-SNE of Latent Space Representations, Colored by Repeater Class.
  • Figure 3: Spearman correlation heatmap between feature space and PCA (left) and t-SNE (right) axes.
  • Figure 4: Latent space traversal for the most impactful features. The change in the top two correlated input features is shown. Repeater probability is shown as a background gradient.