Score-Based Diffusion Models for Bayesian Inference over Star Formation Histories

Sacha Perry-Fagant ( Université de Montréal )

Star formation histories (SFHs) are a key component for understanding galactic evolution. To analyze the evolution of a galaxy, information about its SFHs must be inferred from the galaxy’s photometric data, as SFHs themselves are unobservable. SFHs are poorly constrained by the data so recovering them from observable data is an ill-posed inverse problem. For this reason, SFHs are often modeled using simple parametric models. To generate more realistic SFHs, we train a score-based diffusion model to act as a prior over possible SFHs. Diffusion models employ an annealed sampling technique and have proven to be useful for Bayesian posterior inference. The training data for the model is generated from numerical simulations as, in contrast to real data, the ground truth SFH data is available. Using simulated data also allows the metallicity history to be included in the posterior sampling, which is a key ingredient in simulating a galaxy’s composite stellar population (CSP) and subsequently its photometric data, which is used in posterior inference. Once trained on simulations, the model can be used to infer SFH and metallicity histories for observed galaxies.