In this talk, I will give an overview of how to do field-level likelihood-free inference with galaxy catalogs.
More specifically, only using phase-space information, I will show how to convert galaxy catalogs into graphs.
I will explain how to use graph neural networks, associated with moment neural networks, to constrain Omega matter.
For the first time, we were able to obtain a model which is robust across 5 different hydrodynamic simulations (Astrid,
IllustrisTNG, SIMBA, Magneticum, and SWIFT-EAGLE), i.e., 5 different sub-grid physics models while trained on
a single one. Together I will explain how to do halo-galaxy connection with machine learning from 2 points of view:
one-point value predictions and dealing with probability density distributions. The model performs really well in both
cases, being the second perfect to recover scatter in the astrophysical relations.