AI searches Fermi’s sky for gamma-ray sources

A 27 October preprint presents 1FDL: 3,821 gamma-ray source candidates extracted by deep learning from Fermi observations spanning 2008–2018. Of these, 816 lack a counterpart in the ten-year 4FGL-DR2 catalogue.[1]

Fermi’s Large Area Telescope detects particles produced when gamma rays interact with its instruments. It reconstructs where the incoming radiation came from using those particle tracks and the spacecraft’s orientation. This creates a sky map with an unusual property: the directional blur changes with the photon’s energy. Low-energy photons locate their origin less precisely than high-energy photons.[2]

The background also has structure. Much of the Galactic glow follows interstellar gas, so an apparent concentration of radiation can belong to extended emission rather than a separate object. Standard Fermi analysis fits a model containing neighbouring sources and diffuse emission as well as the source of interest. A candidate’s surroundings are part of the measurement.[2]

The new pipeline combines six energy bands in a neural segmentation map, then refines positions and filters likely spurious detections. Training used simulated point sources.[1]

The two matches are a retrospective check: FL16Y was first posted on 25 February 2026, months before the AI preprint.[1][3]

FL16Y draws on 16 years of observations through August 2024. The survey also recalculated source positions throughout the list, instead of keeping every older position fixed. More exposure and renewed localisation can change which entries line up with one another. The survey team still identifies the diffuse-emission model as a major limitation near the Galactic plane. Longer observation improves the evidence available for a match; the background model remains part of interpreting it.[3]

Report an error or suggest a correction