2026 workshop program
Inference of Star Formation and Metallicity Histories from Galaxy Spectra with Score-Based Models
The work
Abstract
Star formation histories (SFHs) are a key component for understanding galaxy evolution. Because SFHs are not directly observable, they must be inferred from galaxy spectra. However, the mapping from spectra to SFHs is highly degenerate, making the recovery of SFHs an ill-posed inverse problem. As a result, most approaches rely on simple parametric models that impose strong assumptions about the functional form of the history. To generate more realistic histories, we train a score-based diffusion model to act as a prior over SFHs and metallicity histories (MHs). The model is trained on histories from the TNG50 cosmological simulation, where ground-truth values are available. We perform Bayesian posterior sampling by drawing (SFH, MH) pairs from the diffusion model, generating spectra through stellar population synthesis, and comparing them to observed galaxy spectra. Using mock observations from TNG, the method successfully recovers histories consistent with the ground truth, as verified through posterior coverage tests. The approach also generalizes to out-of-distribution histories from the EAGLE simulations. We apply the method to SDSS galaxy spectra and obtain stellar mass estimates consistent with previous studies. These results demonstrate that diffusion-based priors provide a flexible framework for inferring realistic galaxy formation histories from spectroscopic observations.