The First Deep Learning all-sky List of Fermi-LAT Gamma-ray Sources
Abstract
Aims: We develop a deep-learning-based pipeline to construct a list of γ-ray sources and to quantitatively assess the impact of background uncertainties on source detection and characterization. This work extends and enhances the existing AutoSource-ID algorithm, previously adopted on optical observations, for applications on γ-ray data. Methods: Source detection was performed through image segmentation using a U-Net architecture, while dedicated deep neural networks (DNNs) were employed for source localization, flux estimation, and the classification of detected sources as either genuine or spurious. The training data were generated from Monte Carlo simulations based on 10 years of Fermi-LAT observations. Results: We present the first Fermi-LAT γ-ray all-sky source list derived using a deep-learning-based method (1FDL), containing more than 3800 sources with estimated positions, fluxes, and probability scores that quantify the likelihood of each detection being genuine or spurious. We further evaluated the robustness of the pipeline against variations in the interstellar emission model (IEM) and compared the resulting source list with the 4FGL-DR2 catalog. Conclusion: We find that the our deep-learning-based approach achieves performance comparable to, or better than, the likelihood-based methods employed in the standard Fermi-LAT catalog pipeline when tested against changes in the IEM. The 1FDL source list is broadly consistent with the Fermi-LAT 4FGL-DR2 catalog obtained over the same 10-year observational period, including a few new sources confirmed in the recent Fermi-LAT 16-year source list (FL16Y). These results demonstrate the potential of DNN methods for the detection and characterization of γ-ray sources and provide a foundation for their application in future high-energy astrophysical analyses.
Related
- An All-Sky Search for Moving Gamma-Ray Sources with Fermi-LAT
- A Deep Learning Approach to Improving the Arrival-Direction Accuracy of Gamma-Ray Air-Shower Measurements with the Tibet-III Air-Shower Array
- Deep learning methods for gamma event selection in TAIGA-IACT image analysis in stereo-mode
- Testing deep learning techniques for event reconstruction in pair-production gamma-ray telescopes
- A GeV Gamma-Ray Spectral Excess in the Fermi Blazar 4FGL J0136.5+3906
BibTeX
@misc{bhattacharyya2026first,
title = {{The First Deep Learning all-sky List of Fermi-LAT Gamma-ray Sources}},
author = {Saptashwa Bhattacharyya and Giacomo Principe and Dmitry Malyshev and Roberto Ruiz de Austri and Gabrijela Zaharijas and Sascha Caron and Fiorenzo Stoppa and Guðlaugur Johannesson},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.37287},
eprint = {2610.37287},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.37287},
note = {CC BY 4.0},
}
RIS
TY - UNPB
TI - The First Deep Learning all-sky List of Fermi-LAT Gamma-ray Sources
AU - Saptashwa Bhattacharyya
AU - Giacomo Principe
AU - Dmitry Malyshev
AU - Roberto Ruiz de Austri
AU - Gabrijela Zaharijas
AU - Sascha Caron
AU - Fiorenzo Stoppa
AU - Guðlaugur Johannesson
PY - 2026
DA - 2026/10/27
PB - arXiv
DO - 10.48550/arxiv.2610.37287
UR - https://arxiv.org/abs/2610.37287
AB - Aims: We develop a deep-learning-based pipeline to construct a list of γ-ray sources and to quantitatively assess the impact of background uncertainties on source detection and characterization. This work extends and enhances the existing AutoSource-ID algorithm, previously adopted on optical observations, for applications on γ-ray data. Methods: Source detection was performed through image segmentation using a U-Net architecture, while dedicated deep neural networks (DNNs) were employed for source localization, flux estimation, and the classification of detected sources as either genuine or spurious. The training data were generated from Monte Carlo simulations based on 10 years of Fermi-LAT observations. Results: We present the first Fermi-LAT γ-ray all-sky source list derived using a deep-learning-based method (1FDL), containing more than 3800 sources with estimated positions, fluxes, and probability scores that quantify the likelihood of each detection being genuine or spurious. We further evaluated the robustness of the pipeline against variations in the interstellar emission model (IEM) and compared the resulting source list with the 4FGL-DR2 catalog. Conclusion: We find that the our deep-learning-based approach achieves performance comparable to, or better than, the likelihood-based methods employed in the standard Fermi-LAT catalog pipeline when tested against changes in the IEM. The 1FDL source list is broadly consistent with the Fermi-LAT 4FGL-DR2 catalog obtained over the same 10-year observational period, including a few new sources confirmed in the recent Fermi-LAT 16-year source list (FL16Y). These results demonstrate the potential of DNN methods for the detection and characterization of γ-ray sources and provide a foundation for their application in future high-energy astrophysical analyses.
ER -