Semi-supervised morphological classification of fast radio bursts from the second CHIME/FRB catalogue

Technical Summary: Semi-supervised morphological classification of fast radio bursts from the second CHIME/FRB catalogue

Problem Statement
The classification of Fast Radio Bursts (FRBs) based on spectral morphology is critical for understanding their physical origins and progenitor systems. A central challenge in the field is determining whether all FRB sources repeat or if there are distinct populations of repeating and non-repeating (one-off) bursts. Previous studies utilizing the first CHIME/FRB Data Release (DR1) suggested that repeaters tend to have narrower bandwidths and longer temporal widths than one-off bursts. However, these findings have been complicated by instrumental effects, such as side-lobe detections and beam-forming limitations, which can artificially narrow the observed bandwidth of intrinsically broadband bursts. Furthermore, existing classification methods often rely on visual inspection of waterfall plots or supervised learning on manually defined categories, which introduces subjectivity and limits reproducibility. With the release of the second CHIME/FRB Data Release (DR2), containing a significantly larger sample of unique sources, there is a need for a data-driven, reproducible approach to identify morphological subgroups and assess their relationship to repeatability without relying on pre-defined visual labels.

Methodology
The authors propose a semi-supervised approach utilizing a Convolutional Autoencoder (CAE) to perform unsupervised morphological clustering and supervised repeatability classification simultaneously.

• Data and Simulation: To address the class imbalance between repeating and one-off bursts in the observed data (where repeaters are a minority), the authors constructed a synthetic training set. They fitted multi-dimensional Kernel Density Estimates (KDEs) to the fitburst parameters (temporal width, spectral index, spectral running, scattering timescale) of DR1 bursts, separating them into one-off and repeater populations. They simulated 5,000 bursts (2,500 per class) by sampling from these distributions, incorporating realistic Radio Frequency Interference (RFI) masking and thermal noise to mimic CHIME/FRB observational conditions.
• Convolutional Autoencoder (CAE): The core model is a CAE designed to process frequency-time waterfall plots directly. The architecture consists of an encoder with nine convolutional layers compressing the input into a 256-dimensional latent space, and a decoder with nine convolutional layers reconstructing the image. Unlike Variational Autoencoders (VAEs), the CAE does not enforce a Gaussian distribution on the latent parameters, allowing for more flexible representation learning of burst morphology.
• Semi-Supervised Training: While the CAE is trained to minimize reconstruction loss (unsupervised), a classification "head" (a three-layer fully connected network) is simultaneously trained on the 256-dimensional latent representations to predict binary repeatability (repeater vs. one-off) using the simulated labels.
• Clustering and Dimensionality Reduction: The latent representations of real DR1 and DR2 data were processed through Principal Component Analysis (PCA) to reduce noise, followed by Uniform Manifold Approximation and Projection (UMAP) for 2D visualization. Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN) was then applied to the UMAP embeddings to identify distinct morphological groups without forcing all data points into clusters.

Key Contributions

1. Unsupervised Morphological Discovery: The study demonstrates that a CAE trained on simulated data can recover known morphological classes from real CHIME/FRB data without manual annotation. It identifies distinct subgroups that align with previously visually identified categories (e.g., linear negative drifting repeaters) while also revealing new structural nuances.
2. Identification of Distinct Subgroups: The analysis identifies three physically distinct groups separate from the general FRB population:
   • G7: A group of temporally long, spectrally narrow bursts with downward linear drifting, consistent with the "linear negative drifting repeater" class.
   • G4 and G6: Two subgroups of temporally short, simple broadband bursts that occupy the entire CHIME observing band. G4 represents bursts with significant scattering, while G6 represents bursts with minimal scattering.
3. Continuum vs. Discrete Classes: The paper highlights that while extreme morphological groups exist, the majority of FRBs (including the intermediate groups G1, G2, G3, and G5 in DR1) form a continuous distribution in parameter space. In DR2, these intermediate groups merge into a single broad cluster, suggesting that many apparent discrete classes in smaller datasets may result from limited sampling rather than intrinsic physical separation.
4. Repeatability Classification: The semi-supervised classifier achieves a recall of 86% for repeaters on DR2, comparable to state-of-the-art parameter-based methods, demonstrating that latent features extracted directly from waterfall plots encode repeatability information.

Results

• Morphological Overlap: The study finds substantial overlap between the morphological properties of repeaters and one-off bursts. While repeater-rich clusters (like G7) exhibit distinct statistical properties (narrower bandwidth, longer duration, more negative spectral running), one-off bursts are not confined to a single morphological class.
• Repeater Diversity: The classifier correctly identifies 86% of repeaters but also misclassifies some repeaters as one-offs. Specifically, the study observes that individual repeating sources (e.g., FRB 20190905A) can produce bursts with morphologies ranging from typical narrow-band repeater signatures to broadband, short-duration signatures characteristic of one-off bursts (specifically resembling cluster G6).
• Instrumental Effects: The authors note that instrumental effects, particularly side-lobe detections and beam response distortions, likely contribute to the misclassification of broadband one-off bursts as narrow-band repeaters. Approximately 30% of bursts may be detected in side lobes, potentially suppressing parts of their spectrum and mimicking repeater morphology.
• Scattering Distinction: The separation of broadband bursts into G4 (scattered) and G6 (unscattered) suggests that scattering by plasma screens is a key differentiator within the one-off population, potentially indicating different propagation environments rather than different progenitor types.

Significance and Claims
The paper claims that unsupervised, data-driven methods can objectively recover FRB morphological structures previously defined by visual inspection, validating the physical reality of these categories. However, the authors modestly conclude that burst morphology alone does not provide a definitive indicator of repeatability.

While certain morphological groups are statistically associated with repeaters (e.g., G7) or one-offs (e.g., G4, G6), the significant overlap in parameter space and the intra-source variability of repeating bursts suggest that repeaters and one-offs may share a common emission mechanism or that the observed continuum reflects a combination of intrinsic diversity and instrumental biases. The study emphasizes that the apparent separation of FRB populations is likely a continuum rather than discrete classes, and that future analysis using higher-resolution baseband data is required to disentangle intrinsic morphological diversity from instrumental artifacts like side-lobe effects and sub-millisecond sub-burst structure.