Inferring Galactic Parameters from Chemical Abundances: A Multi-Star Approach

TL;DR

This paper presents a Bayesian method using chemical abundances and ages of stars, combined with neural network acceleration, to accurately infer key galactic parameters, demonstrating high precision with simulated data.

Contribution

The novel approach integrates chemical evolution modeling with neural network surrogates and Bayesian inference to precisely determine galactic parameters from stellar abundances.

Findings

Achieved sub-percent accuracy in parameter inference using simulated data.

Metal ratios provide the strongest constraints among chemical abundances.

Model errors and yield assumptions significantly impact parameter estimates.

Abstract

Constraining parameters such as the initial mass function high-mass slope and the frequency of type Ia supernovae is of critical importance in the ongoing quest to understand galactic physics and create realistic hydrodynamical simulations. In this paper, we demonstrate a method to precisely determine these using individual chemical abundances from a large set of stars, coupled with some estimate of their ages. Inference is performed via the simple chemical evolution model Chempy in a Bayesian framework, marginalizing over each star's specific interstellar medium parameters, including an element-specific `model error' parameter to account for inadequacies in our model. Hamiltonian Monte Carlo (HMC) methods are used to sample the posterior function, made possible by replacing Chempy with a trained neural network at negligible error. The approach is tested using data from both Chempy and the IllustrisTNG simulation, showing sub-percent agreement between inferred and true parameters using data from up to 1600 individual stellar abundances. For IllustrisTNG, strongest constraints are obtained from metal ratios, competitive with those from other methods including star counts. Analysis using a different set of nucleosynthetic yields shows that incorrectly assumed yield models can give non-negligible bias in the derived parameters; this is reduced by our model errors, which further show how well the yield tables match data. We also find a significant bias from analyzing only a small set of stars, as is often done in current analyses. The method can be easily applied to observational data, giving tight bounds on key galactic parameters from chemical abundances alone.