2024 workshop program
Inferring Star formation histories with neural density estimation and cosmological simulations
The work
Abstract
We present a simulation-based inference approach for reconstructing the Star Formation History (SFH) of galaxies using photometric data, informed by hydrodynamical simulations, moving beyond the traditional use of analytical priors. Our method bridges the gap between cosmological simulations and observation, leveraging Neural Posterior Estimation to perform Bayesian inference with a physically motivated prior. This allows to align the inferential process more closely with the underlying physics of galaxy formation and its interactions with galaxies environment.
Our approach combines a non parametric description of the SFH and Normalizing flows ability in modeling complex probability distributions, allowing for an informed but flexible inference of the star formation history in galaxies. The cosmological hydrodynamical simulation provides a robust framework encoding our understanding of large scale structures formation, from which we derive our physically motivated priors. This is a significant departure from Spectral Energy Distribution (SED) fitting, offering a method that is both theoretically grounded and practically effective.
We demonstrate the performances on simulations as well as real observations and explore potential biases it induces.