Searching for dark matter subhalos in the Fermi-LAT catalog with Bayesian neural networks

Machine learning techniques are powerful tools to tackle diverse tasks in current astroparticle physics research. For example, Bayesian neural networks provide robust classifiers with reliable uncertainty estimates, and are particularly well suited for classification problems that are based on comparatively small and imbalanced data sets, such as the gamma-ray sources detected by Fermi-Large Area Telescope (LAT).
About one third of the gamma-ray sources collected in the most recent catalogs remain currently unidentified. Intriguingly, some of these could be exotic objects such as dark subhalos, which are overdensities in dark matter halos predicted to form by cosmological N-body simulations. If they exist in the Milky Way, they could be detected as gamma-ray point sources due to the annihilation or decay of dark matter particles into Standard Model final states.
In this talk I will discuss our recent work* in which, after training on realistic simulations, we use Bayesian neural networks to identify candidate dark matter subhalos among unidentified gamma-ray sources in Fermi-LAT catalogs. Our novel framework allows us to derive conservative bounds on the dark matter annihilation cross section, by excluding unidentified sources classified as astrophysical-like.