Researchers used a deep‑learning pipeline called 1FDL to scan ten years of Fermi‑LAT data. The method first segments sky maps with a Multi‑Input U‑Net, then locates candidate sources with Laplacian‑of‑Gaussian blobs. Dedicated neural networks refine the positions, estimate fluxes above 1 GeV, and classify each candidate as genuine or spurious.

The resulting list has 3 821 entries. More than 98 % of bright, high‑latitude sources (Signif_Avg > 20) match those in the 4FGL‑DR2 catalogue, and 75 % of all sources with Signif_Avg > 10 are recovered across the sky. For high‑latitude sources the effective precision and recall stay above 90 % for fluxes above 2 × 10⁻¹⁰ ph cm⁻² s⁻¹.

The pipeline’s sensitivity, defined as the flux at which 90 % of simulated sources are found, is about 2.3 × 10⁻¹⁰ ph cm⁻² s⁻¹ for |b| > 20°, matching the 4FGL completeness threshold.

Two bright candidates in the catalogue—1FDL J1647.9−1209 and 1FDL J0535.7−6610—were later confirmed in the 16‑year Fermi‑LAT list, showing that the method can uncover previously missing sources.

Limitations remain. The pipeline was trained only on simulated point‑like emitters, so it often misses bright extended sources and struggles with close pairs separated by less than 0.1°, especially near the Galactic plane. A post‑processing classifier reduces the false‑positive rate to about 7 %. The analysis used only front‑converted LAT events between 300 MeV and 1 TeV, excluding back‑converted events with poorer angular resolution.