Classification and physical characteristics analysis of Fermi-GBM Gamma-ray bursts based on Deep-learning

TL;DR

This paper introduces a deep learning-based method using CNNs to classify Fermi-GBM Gamma-ray bursts, effectively resolving the overlap issue in traditional duration-based classification and revealing distinct physical properties of the categories.

Contribution

The study presents a novel CNN-based classification approach that refutes the intermediate GRB class and links physical properties with burst types, outperforming traditional methods.

Findings

Successfully classified overlapping GRBs into two categories.

Discovered two distinct clusters corresponding to S-type and L-type GRBs.

Identified spectral differences and associations with kilonovae and supernovae.

Abstract

The classification of Gamma-Ray Bursts has long been an unresolved problem. Early long and short burst classification based on duration is not convincing due to the significant overlap in duration plot, which leads to different views on the classification results. We propose a new classification method based on Convolutional Neural Networks and adopt a sample including 3774 GRBs observed by Fermi-GBM to address the $T_\text{90}$ overlap problem. By using count maps that incorporate both temporal and spectral features as inputs, we successfully classify 593 overlapping events into two distinct categories, thereby refuting the existence of an intermediate GRB class. Additionally, we apply the optimal model to extract features from the count maps and visualized the extracted GRB features using the t-SNE algorithm, discovering two distinct clusters corresponding to S-type and L-type GRBs. To further investigate the physical properties of these two types of bursts, we conduct a time-integrated spectral analysis and discovered significant differences in their spectral characteristics. The analysis also show that most GRBs associated with kilonovae belong to the S-type, while those associated with supernovae are predominantly L-type, with few exceptions. Additionally, the duration characteristics of short bursts with extended emission suggest that they may manifest as either L-type or S-type GRBs. Compared to traditional classification methods (Amati and EHD methods), the new approach demonstrates significant advantages in classification accuracy and robustness without relying on redshift observations. The deep learning classification strategy proposed in this paper provides a more reliable tool for future GRB research.