Robust field-level likelihood-free inference with galaxies

Paper highlights:

• Uses graph neural networks to perform field-level, likelihood-free cosmological inference directly from galaxy catalogs in the CAMELS simulations.
• Achieves ~12% precision on Ωm using only galaxy positions and velocities in small volumes, while remaining robust across multiple hydrodynamical codes and feedback models.
• Demonstrates strong extrapolation across a wide parameter space, indicating the networks have learned a galaxy-formation–independent physical relation.

You can read the paper here.