Hunting dark matter subhalos in the gamma ray sky with machine learning

A large fraction of the gamma-ray sources identified by Fermi-LAT remains currently unassociated, posing a classification challenge and potentially hiding a population of dark matter subhalos. A key modeling difficulty is the significant shift in the energy spectrum distribution between unassociated and associated sources.

In the first part of this talk, I will present a new statistical search that addresses this shift using a generative machine learning model. This model treats unassociated sources as a mixture of known populations and a potential dark matter signal, explicitly accounting for data shifts. I will show how this model can be used both for source classification and dark matter constraints.

In the second part, I will lay the groundwork for identifying dark matter subhalos among unresolved gamma-ray sources using machine learning techniques. I will describe a method to count the number of unresolved gamma-ray sources as a function of their flux, effectively reconstructing the source-count distribution below Fermi-LAT nominal sensitivity.