Unveiling the Spectral Morphological Division of Fast Radio Bursts with CHIME/FRB Catalog 2

TL;DR

This study analyzes CHIME/FRB Catalog 2 with unsupervised learning (UMAP) and density-based clustering (HDBSCAN) on eight spectral and temporal features for 4527 FRBs, revealing a robust bimodal division into Repeater-like and Nonrepeater-like populations. The spectral morphology parameter $r$ emerges as the dominant discriminator, supporting the view that narrowband emission is intrinsic to repeating FRBs. A stable subclass of atypical repeaters (~6% of repeaters) resides in the Nonrepeater-like space and exhibits broadband, high-energy properties, suggesting a physical connection or continuum between the two classes. The results imply that apparently non-repeating FRBs may reflect observational incompleteness and that repeating and non-repeating FRBs are not strictly separate populations, highlighting multiple emission modes within repeaters and a need for larger, multi-parameter surveys to fully map FRB diversity.

Abstract

Fast radio bursts (FRBs) are commonly divided into repeating and apparently non-repeating sources, but whether these represent distinct physical populations remains uncertain. In this work, we apply an unsupervised machine learning methods combining Uniform Manifold Approximation and Projection (UMAP) with density-based clustering to analyze CHIME/FRB Catalog 2. We find that FRBs remain primarily separated into two clusters in the multi-dimensional parameter space, with a recall of 0.94 for known repeaters, indicating strong robustness. Consistent with Catalog 1 analyses, we confirm that the spectral morphology parameter, specifically spectral running remains the key discriminator between the two populations, indicating that narrowband emission is an intrinsic and persistent property of repeating FRBs. With the enlarged Catalog 2 sample, we further identify a stable subclass of atypical repeaters (about $6\%$ of repeating bursts) that are broadband, shorter in duration, and more luminous, resembling non-repeating bursts. The Nonrepeater-like cluster also shows higher inferred energies and dispersion measures, consistent with a scenario in which apparently non-repeating FRBs may result from observational incompleteness, with low-energy repeating bursts remaining undetected. Our results provide new statistical evidence for a physical connection between repeating and non-repeating FRBs.

Figures (6)

• Figure 1: The embedding space of the UMAP dimension reduction results of the Catalog 2. The blue dots represent the non-repeaters and the red dots are repeaters. The blue and red contours indicate the clusters identified by HDBSCAN, corresponding to the Nonrepeater-like cluster and Repeater-like cluster, respectively. The surrounding panels display the dynamic spectra for representative FRBs selected from both clusters.
• Figure 2: Confusion matrix of the clustering classification.
• Figure 3: Distribution of the proportion of atypical repeating bursts within the repeater sources.
• Figure 4: SHAP values for predictions in FRB classification. Features are ranked by overall importance, with the horizontal axis indicating the SHAP value, reflecting each feature’s impact on the model output.
• Figure 5: Parameter distribution comparison between the Repeater-like and Nonrepeater-like clusters. The $p_{\rm AD}$ and $s_{\rm AD}$ values in each panel denote the $p$-value and test statistic of the AD test, respectively, indicating the statistical significance of the distributional differences between the two clusters.