Hostname: page-component-7b7567c8c-tnkh4 Total loading time: 0 Render date: 2026-11-17T15:46:47.012Z Has data issue: false hasContentIssue false

Unsupervised Classification of Gamma-ray Bursts from Blazars (GRBBLs) with Machine Learning

Published online by Cambridge University Press:  05 October 2026

Matteo Cerruti*
Affiliation:
Université Paris Cité, CNRS, Astroparticule et Cosmologie, F-75013 Paris, France
Rights & Permissions [Opens in a new window]

Abstract

Blazars dominate the extragalactic $\gamma$-ray sky and show pronounced flares. Using public Fermi-LAT light curves for 732 blazars with secure redshifts, I implement an automated pipeline to identify and characterize $\gamma$-ray bursts from blazars (GRBBLs). Each event is modeled with an exponential rise/decay profile, and spectral variability is quantified via a constant fit. From 679 high-quality GRBBLs, I apply extreme deconvolution for unsupervised classification. The GRBBL population is remarkably homogeneous; the most robust split is in achromatic vs. chromatic events, with significant overlap. Removing spectral information yields a luminosity-driven classification in type-1 and type-2 GRBBLs, although this classification is not identified in all tests. This study establishes GRBBL population studies as a tool to study blazars. As a by-product of this project I identify a correlation between peak luminosity and timescales in GRBBLs.

Information

Type
Contributed Paper
Copyright
© The Author(s), 2026. Published by Cambridge University Press on behalf of The International Astronomical Union
Figure 0

Figure 1.A GRBBL identified in 4FGL J0108.6+0134, from Cerruti (2025). Top: Luminosity vs time (the vertical green lines indicate t0$t_0$ (the central one), t0−2τr$t_0 - 2\tau_r$ (on the left), and t0+2τd$t_0 + 2\tau_d$ (on the right); the red line represents the fit function); and fit residuals vs. time; Bottom: Photon index vs. time (the horizontal green lines show the best-fit value of Γ$\Gamma$ plus or minus its uncertainty; the horizontal red lines show the same but for Γ4LAC$\Gamma_{4LAC}$).

Figure 1

Figure 2.Results for the Extreme Deconvolution algorithm run on eight parameters. Top: correlation plot for τd$\tau_d$ vs τr$\tau_r$ (error bars have been removed for clarity) color-coded by the most likely cluster for k=2$k=2$; same as before, but for the Γ$\Gamma$ vs χΓ2~$\tilde{\chi^2_\Gamma}$ correlation. Figure taken from Cerruti (2025).