Fermi Gamma-Ray Burst Classification Based on a Convolutional Autoencoder

Abstract

Traditionally, gamma-ray bursts (GRBs) are classified into long and short bursts based on the bimodal distribution of their duration. However, this phenomenological dichotomy has been challenged by recent observations, and current classification methods often rely on manual pre-analysis of light curves. In this paper, we propose an end-to-end feature learning and classification method based on a convolutional autoencoder (CAE), which enables automatic feature extraction and unsupervised classification directly from raw GRB count images. Using 3819 GRB samples observed by the Fermi Gamma-ray Burst Monitor (GBM) from July 2008 to January 2024, we construct count images with a resolution of 512 × 512 pixels that integrate temporal and spectral information. The CAE model is built upon a pre-trained ResNet-18 encoder with a convolutional block attention module (CBAM), learning high-dimensional feature vectors, which are then visualized using the UMAP algorithm. The results show that GRBs do not exhibit a clear subclass separation structure and are essentially continuously distributed. Key physical parameters such as T 90 , E p , S γ , and F p show clear evolutionary gradients. Short bursts tend to cluster in specific regions, supporting the idea that GRBs may indeed differ in their origins, but without a clear boundary in their physical properties.