Neural Stellar Population Synthesis Emulator for the DESI PROVABGS

TL;DR

This paper introduces a neural emulator for stellar population synthesis models to efficiently infer galaxy properties in the DESI PROVABGS survey, achieving high accuracy and 100-fold speedup.

Contribution

It develops and validates a neural emulator for SPS models, enabling scalable Bayesian analysis of millions of galaxy observations.

Findings

Emulator achieves <1% error compared to original SPS model.

Emulator is 100 times faster than traditional methods.

Posteriors from the emulator closely match those from the original model.

Abstract

The Probabilistic Value-Added Bright Galaxy Survey (PROVABGS) catalog will provide the posterior distributions of physical properties of $>10$ million DESI Bright Galaxy Survey (BGS) galaxies. Each posterior distribution will be inferred from joint Bayesian modeling of observed photometry and spectroscopy using Markov Chain Monte Carlo sampling and the [arXiv:2202.01809] stellar population synthesis (SPS) model. To make this computationally feasible, PROVABGS will use a neural emulator for the SPS model to accelerate the posterior inference. In this work, we present how we construct the emulator using the [arXiv:1911.11778] approach and verify that it can be used to accurately infer galaxy properties. We confirm that the emulator is in excellent agreement with the original SPS model with $\ll 1\%$ error and is $100\times$ faster. In addition, we demonstrate that the posteriors of galaxy properties derived using the emulator are also in excellent agreement with those inferred using the original model. The neural emulator presented in this work is essential in bypassing the computational challenge posed in constructing the PROVABGS catalog. Furthermore, it demonstrates the advantages of emulation for scaling sophisticated analyses to millions of galaxies.