Unsupervised machine learning classification of Fermi gamma-ray bursts using spectral parameters

A thorough analysis of 2297 gamma-ray bursts (GRBs) in the Fermi catalogue is performed by using unsupervised machine learning algorithms in this paper. In our analysis, for two spectral parameter samples, namely for the peak-flux and time-integrated spectral fits, two dimensionality reduction algorithms, t-distributed stochastic neighbour embedding (t-SNE), and uniform manifold approximation and projection (UMAP), are used to generate four embedding maps; further, K-means algorithm is used for searching for the optimal clustering on the four maps. Our results show that Fermi GRBs can be well separated into two groups. For the time-integrated spectral parameters, both UMAP and t-SNE algorithms classify 372 bursts as short GRBs and 1925 bursts as long GRBs, and 384 bursts as short GRBs and 1913 bursts as long GRBs for the peak-time spectral parameters. This new classification method differs from traditional long and short classifications because it is not based on duration. In addition, it is found that the classification results of 11 GRBs are inconsistent between the integrated and peak-time spectral samples. GRB200826A is the first confirmed short GRB of collapsar origin, and the physical origins of these GRBs may be similar to it.