Title: Hierarchical Bayesian inference of photometric redshifts with stellar population synthesis models

Abstract: We present a Bayesian hierarchical framework to analyze photometric galaxy survey data with stellar population synthesis (SPS) models. Our method couples robust modeling of spectral energy distributions with a population model and a noise model to characterize the statistical properties of the galaxy populations and real observations, respectively. By self-consistently inferring all model parameters, from high-level hyper-parameters to SPS parameters of individual galaxies, one can separate sources of bias and uncertainty in the data.We demonstrate the strengths and flexibility of this approach by deriving accurate photometric redshifts for a sample of spectroscopically-confirmed galaxies in the COSMOS field, all with 26-band photometry and spectroscopic redshifts. We achieve a performance competitive with publicly-released photometric redshift catalogs based on the same data. Prior to this work, this approach was computationally intractable in practice due to the heavy computational load of SPS model calls; we overcome this challenge using with neural emulators. We find that the largest photometric residuals are associated with poor calibration for emission line luminosities and thus build a framework to mitigate these effects. This combination of physics-based modeling accelerated with machine learning paves the path towards meeting the stringent requirements on the accuracy of photometric redshift estimation imposed by upcoming cosmological surveys. The approach also has the potential to create new links between cosmology and galaxy evolution through the analysis of photometric datasets.