Robust Field-level Likelihood-free Inference with Galaxies

We train graph neural networks to perform field-level likelihood-free inference using galaxy catalogs from state-of-the-art hydrodynamic simulations of the CAMELS project. Our models are rotational, translational, and permutation invariant do not impose any cut on scale. From that only contain $3$D positions radial velocities $\sim 1, 000$ galaxies in tiny $(25~h^{-1}{\rm Mpc})^3$ volumes our can infer value $\Omega_{\rm m}$ with approximately $12$ % precision. More importantly, by testing thousands simulations, each having a different efficiency supernova AGN feedback, run five codes subgrid - IllustrisTNG, SIMBA, Astrid, Magneticum, SWIFT-EAGLE -, we find robust changes astrophysics, physics, subhalo/galaxy finder. Furthermore, test $1,024$ cover vast region parameter space variations $5$ cosmological $23$ astrophysical parameters finding model extrapolates really well. results indicate key building is use both velocities, suggesting network have likely learned an underlying physical relation does depend formation valid scales larger than $\sim10~h^{-1}{\rm kpc}$.