Unsupervised machine learning classification of gamma-ray bursts based on the rest-frame prompt emission parameters

¹ College of Physics and Electronic Information Engineering, Guilin University of Technology, Guilin 541004, China
² School of Physics and Astronomy, Sun Yat-Sen University, Zhuhai 519082, China
³ Department of Space Sciences and Astronomy, Hebei Normal University, Shijiazhuang 050024, China
⁴ Key Laboratory of Low-dimensional Structural Physics and Application, Education Department of Guangxi Zhuang Autonomous Region, Guilin 541004, China

^⋆ Corresponding author: fwzhang@pmo.ac.cn

Received: 5 July 2025
Accepted: 7 September 2025

Abstract

Gamma-ray bursts (GRBs) are generally believed to originate from two distinct progenitors, namely, compact binary mergers and massive collapsars. Traditional methods, along with a number of recent machine learning-based classification schemes, predominantly rely on observer-frame physical parameters, which are significantly affected by the redshift effects and might not accurately represent the intrinsic properties of GRBs. In particular, the progenitors could typically only be determined on the basis of a successful detection of the multiband long-term afterglow, which could easily cost days of devoted effort from multiple global observational utilities. In this work, we applied the unsupervised machine learning (ML) algorithms t-SNE and UMAP to perform the GRB classification based on rest-frame prompt emission parameters. The map results of both t-SNE and UMAP reveal a clear division of these GRBs into two clusters, denoted as GRBs-I and GRBs-II. We find that all supernova-associated GRBs, including the atypical short-duration burst GRB 200826A (now recognized as collapsar-origin), consistently fall within the GRBs-II category. Conversely, all kilonova-associated GRBs (except for two controversial events) are classified as GRBs-I, including the peculiar long-duration burst GRB 060614 originating from a merger event. In another words, this clear ML separation of two types of GRBs based only on prompt properties could correctly predict the results of progenitors without follow-up afterglow properties. A comparative analysis with conventional classification methods using T₉₀ and E_p, z–E_iso correlation demonstrates that our machine learning approach provides superior discriminative power, particularly with respect to resolving ambiguous cases of hybrid GRBs.