The formation and evolution of structures in the Universe is expected to be affected by the nature and properties of dark matter. For this reason, we can made use of cosmological and astrophysical observations to learn about dark matter. Unfortunately, there are multiple intrinsic complications associated to the search of those signatures: 1) numerical simulations are usually needed to capture the non-linear physics involved on the relevant scales, 2) uncertain astrophysical processes can mimic the dark matter signal, and 3) the lack of an optimal quantity where the expected signal is maximum complicates the search. In this seminar I will present a collection of ideas on how we can made use of deep learning to approach these problems. I will first show how neural networks can to focus on the relevant physics while marginalizing over astrophysical processes, even if the optimal property is unknown. Next, I will present preliminary results obtained when using these techniques to constrain the mass of warm dark matter.