Representation learning for fast radio burst dynamic spectra

Dirk Kuiper, Gabriella Contardo, Daniela Huppenkothen, J. W. T. Hessels

Abstract

Fast radio bursts (FRBs) are millisecond-duration radio transients of extragalactic origin, with diverse time-frequency patterns and emission properties that require explanation. With one possible exception, FRBs are detected only in the radio, so analyzing their dynamic spectra is therefore crucial to disentangling the physical processes governing their generation and propagation. Furthermore, comparing FRB morphologies provides insights into possible differences among their progenitors and environments. This study applies unsupervised learning and deep learning techniques to investigate FRB dynamic spectra, focusing on two approaches: Principal Component Analysis (PCA) and a Convolutional Autoencoder (CAE) enhanced by an Information-Ordered Bottleneck (IOB) layer. PCA served as a computationally efficient baseline, capturing broad trends, identifying outliers, and providing valuable insights into large datasets. However, its linear nature limited its ability to reconstruct complex FRB structures. In contrast, the IOB-augmented CAE excelled at capturing intricate features, with high reconstruction accuracy and effective denoising at modest signal-to-noise ratios. The IOB layer's ability to prioritize relevant features enabled efficient data compression, preserving key morphological characteristics with minimal latent variables. When applied to real FRBs from CHIME, the IOB-CAE generalized effectively, revealing a latent space that highlighted the continuum of FRB morphologies and the potential for distinguishing intrinsic differences between burst types. This framework demonstrates that while FRBs may not naturally cluster into discrete groups, advanced representation learning techniques can uncover meaningful structures, offering new insights into the diversity and origins of these bursts.

Introduction

Fast Radio Bursts (FRBs) are intense, millisecond-long flashes of radio waves originating from distant galaxies. Despite thousands of detections, their exact physical origins remain one of the most compelling mysteries in modern astrophysics. The key to unlocking these origins lies in their dynamic spectra—often visualized as "waterfall plots"—which show how the burst's intensity evolves across both time and frequency.

These spectra are high-dimensional and incredibly diverse. Some bursts are broad and simple, while others exhibit complex sub-burst structures or a downward frequency drift known as the "sad trombone" effect. Traditionally, astronomers have relied on manual classification or simple linear models to categorize these bursts. However, such methods struggle to capture the non-linear, intricate structures of the most complex signals. This research seeks an objective, machine-learning-driven framework to represent these bursts in a low-dimensional space, uncovering hidden patterns and relationships between different burst morphologies.

Datasets and Morphologies

To train and validate the representation learning models, the study utilized two primary data sources. First, the researchers developed FRBakery, a simulation tool used to generate 5,000 synthetic bursts. These bursts were categorized into five distinct morphological types: Simple Narrow, Simple Broad, Scattered, Complex, and Drifting. These simulations provided a "ground truth" for the models to learn the underlying physics of signal dispersion and scattering.

Examples of simulated FRB morphologies across five categories: Simple Broad, Simple Narrow, Scattered, Complex, and Drifting.

Second, the models were applied to real-world observational data from the CHIME/FRB Catalog, consisting of 140 high-time-resolution bursts. This step was critical to ensure that the features learned from synthetic data could generalize to the noisy, complex reality of actual telescope observations.

Methods

The study compared two primary approaches for representation learning:

  1. Principal Component Analysis (PCA): A classic linear method that identifies the directions of maximum variance in the data. While computationally efficient and useful for identifying broad trends or instrumental outliers, PCA often fails to reconstruct the fine-grained, non-linear details of complex astrophysical signals.
  2. IOB-CAE (Convolutional Autoencoder with Information-Ordered Bottleneck): This novel architecture combines the spatial pattern recognition of convolutional layers with a specialized bottleneck. The Information-Ordered Bottleneck (IOB) forces the network to prioritize information: the first latent variable captures the most critical feature, with subsequent variables capturing increasingly fine details. This makes the resulting compressed representation highly interpretable.
Architecture of the IOB-augmented Convolutional Autoencoder (CAE) used for representation learning.

Key Results and Insights

The Morphological Continuum

A significant finding of this work is that FRBs do not appear to form distinct, isolated clusters in the latent space. Instead, they exist on a morphological continuum. This suggests that the physical processes or environments creating these bursts likely share overlapping characteristics, rather than belonging to fundamentally different, discrete "types."

Reconstruction and Denoising

The IOB-CAE demonstrated superior performance in reconstructing complex signals compared to PCA. While PCA produced blurry results for multi-component bursts, the IOB-CAE stabilized its reconstruction quality with just 6–8 latent variables. Furthermore, the model acted as a powerful unsupervised denoiser; by forcing the signal through a narrow bottleneck, the network naturally discarded random background noise, preserving only the structural astrophysical signal.

Outlier Detection and Generalization

Using PCA, the team identified nine specific bursts in the CHIME dataset with unusual characteristics. Some were revealed to be "channelization artifacts"—spurious signals introduced by the telescope's hardware—demonstrating the model's utility in data quality control. Crucially, the IOB-CAE successfully mapped real CHIME bursts into the same latent space as the simulated ones, proving that the learned features are physically meaningful.

Conclusion and Future Work

This framework demonstrates that advanced representation learning can uncover meaningful structures in FRB data without the need for manual labels. While the current study used 140 real bursts, the latent space will become even better defined as the CHIME catalog grows. Future work will involve incorporating more complex physics into the FRBakery simulator, such as polarization and micro-structure, and exploring alternative loss functions to better preserve sharp, fine-grained spectral details.

Summary updated

Abstract

Fast radio bursts (FRBs) are millisecond-duration radio transients of extragalactic origin, with diverse time-frequency patterns and emission properties that require explanation. With one possible exception, FRBs are detected only in the radio, so analyzing their dynamic spectra is therefore crucial to disentangling the physical processes governing their generation and propagation. Furthermore, comparing FRB morphologies provides insights into possible differences among their progenitors and environments. This study applies unsupervised learning and deep learning techniques to investigate FRB dynamic spectra, focusing on two approaches: Principal Component Analysis (PCA) and a Convolutional Autoencoder (CAE) enhanced by an Information-Ordered Bottleneck (IOB) layer. PCA served as a computationally efficient baseline, capturing broad trends, identifying outliers, and providing valuable insights into large datasets. However, its linear nature limited its ability to reconstruct complex FRB structures. In contrast, the IOB-augmented CAE excelled at capturing intricate features, with high reconstruction accuracy and effective denoising at modest signal-to-noise ratios. The IOB layer's ability to prioritize relevant features enabled efficient data compression, preserving key morphological characteristics with minimal latent variables. When applied to real FRBs from CHIME, the IOB-CAE generalized effectively, revealing a latent space that highlighted the continuum of FRB morphologies and the potential for distinguishing intrinsic differences between burst types. This framework demonstrates that while FRBs may not naturally cluster into discrete groups, advanced representation learning techniques can uncover meaningful structures, offering new insights into the diversity and origins of these bursts.
BibTeX
@inproceedings{kuiper2025representation,
  author = {Dirk Kuiper and Gabriella Contardo and Daniela Huppenkothen and J. W. T. Hessels},
  title = {{Representation learning for fast radio burst dynamic spectra}},
  booktitle = {Monthly},
  year = {2025},
  doi = {10.1093/mnras/staf306},
}
Figure