The Fermi LAT point source catalog contains 10 years of observational data between 50 MeV to 1 TeV. It contains 5064 point sources mostly consisting of BLLs (1131) and FSRQs (694), while pulsars (239) are the most numerous Galactic population. However, a quarter of detected sources remains unclassified and might hide new source classes. The classification is difficult due to bright, diffuse emission from our own galaxy.Recently a machine learning methods were developed for the first time to localize and to classify point sources in the catalog, with performance comparable to that of traditional techniques. Synthetic yearly catalogs are simulated to produce 10 yearly $\gamma$-ray images from 2008 to 2018 in 6 energy bins of the sources. The yearly images provide the network with time variability information of the point sources. The time variable images are fed to the new neural network together with the location in the sky of the point source.
The network then separates the sources into distinct classes. The addition of time dependency and location data should increase the number of classifiable sources compared to the previous network from 3 to 5 (BLLacs, FSRQs, PSRs, PWN+SNR+SPPs, and Fakes), as well as an increase in classification accuracy.
machine learning; Fermi LAT; 4FGL catalog; neural network; source classification