Multivariate Time Series Classification of Fermi-Detected Gamma-Ray Transients Using Convolutional-Recurrent Neural Networks

TL;DR

This work tackles the rapid, robust classification of Fermi-GBM gamma-ray transients in the era of multi-messenger astronomy. It introduces two convolutional-recurrent neural networks (R-ConvNet and MI-DCL) that ingest multivariate time-series light curves from GBM TTE data across two energy bands and multiple time resolutions to classify GRBs, TGFs, SGRs, and SFLAREs, with an explicit UNCERT class for unknowns. The models achieve about 93% overall accuracy and identify ~1-2% as outliers, while enabling fast per-event inference suitable for onboard or automated pipelines. The results demonstrate the value of multi-resolution temporal features and uncertainty handling for rapid transient identification and discovery of new phenomena.

Abstract

Fermi Gamma-ray Space Telescope has detected a diverse range of gamma-ray transients since its launch in 2008. Over the years, Fermi has accumulated an extensive public archive of transient events. Traditional classification methods for these events typically rely on fixed thresholds, localisation accuracy, and characteristic light curve features. However, in the current era of time-critical, multi-wavelength, and multi-messenger astronomy, rapid and reliable classification is essential to enable timely follow-up and coordinated observations. In this work, we develop and present two deep learning-based classifiers that integrate convolutional and recurrent neural network architectures. Using multivariate time-series inputs derived from Fermi-GBM data, our models are trained to distinguish among four classes of gamma-ray transients: Gamma-Ray Bursts (GRBs), Terrestrial Gamma-ray Flashes (TGFs), Solar Flares (SFLAREs), and Soft Gamma Repeaters (SGRs). Furthermore, the models are designed to flag events that do not conform to any of these categories, providing a pathway for identifying potentially new or rare transient types. Training was conducted using a carefully curated subset of high-confidence Fermi events. The resulting models achieve an overall classification accuracy of 93%, and identify approximately 2.5% of the triggers as outliers of unknown origin. When applied to Fermi events with uncertain classifications, our models assign 60% of them to the TGF category with over 60% confidence. These results demonstrate that incorporating deep learning-based classification into onboard or automated data pipelines can significantly enhance transient identification, minimize misclassification, and improve the discovery potential of new phenomena in future high-energy astrophysics missions.

Figures (11)

• Figure 1: Examples of the lightcurves of different transient events, shown with different time bin sizes. (a) GRB bn100820373 with a bin size of 0.256 s in the 100–900 keV energy range, displaying a characteristic burst profile with rapid variability. (b) TGF bn120827282 with a bin size of 0.004 s in the 100–900 keV energy range, showing an extremely short-duration peak typical of terrestrial gamma– ray flashes. (c) SGR bn210911652 with a bin size of 0.016 s in the 10–100 keV energy range, exhibiting multiple short spikes, a signature of magnetar bursts. (d) SFLARE bn110924207 with a bin size of 4.096 s in the 10–100 keV energy range, highlighting a gradual rise and decay structure often seen in solar flares.
• Figure 2: The background-subtracted light curves used as model inputs are shown above. Each sample comprises 14 light curves from two energy ranges (10–100 keV and 100–900 keV) and seven time-bin resolutions. When the algorithm fails to clearly separate signal from background, an empty light curve is passed to the model.
• Figure 3: Schematic illustration of the data processing pipeline, showing the complete workflow from data extraction to background removal and generation of clean data arrays.
• Figure 4: The schematic diagram of the proposed R-ConvNet model illustrates the detailed architecture, showing how the input data are processed through successive layers to learn underlying spatial and temporal features.
• Figure 5: The schematic diagram of the proposed MI-DCL model illustrates the detailed architecture and how the input data are processed through successive layers to learn underlying spatial and temporal features.