Speaker
Description
Fast radio bursts (FRBs) are bright, millisecond-duration radio transients with diverse morphologies and uncertain origins. As detection rates accelerate, manual classification becomes impractical. In this talk, I present a new framework for exploring FRB time-frequency structures using unsupervised machine learning techniques. We apply Principal Component Analysis (PCA) and a Convolutional Autoencoder (CAE) enhanced with an Information-Ordered Bottleneck (IOB) to both simulated and real FRB dynamic spectra, including high-time-resolution data from CHIME/FRB. PCA provides a fast, interpretable baseline for identifying broad trends and outliers, while the IOB-CAE excels at capturing complex, non-linear burst features with high reconstruction fidelity—even at low signal-to-noise. Using a newly developed FRB simulation tool, FRBakery, we evaluate these methods across diverse morphologies, demonstrating the potential of latent representations to reveal a continuum of FRB types and uncover subtle structure in large datasets. Our approach enables efficient, scalable analysis of FRB populations and provides a foundation for future classification efforts in the era of data-intensive radio astronomy.
| Talk category | NOVA Network 3 |
|---|---|
| Preference for a talk or poster | Talk |
| Talk preference for PhD students | 1st year |
