Data-Driven Stellar Models

TL;DR

This paper presents a neural network-based data-driven model that accurately predicts stellar photometry from spectroscopic parameters and parallaxes, simultaneously constraining dust reddening and improving stellar property inference.

Contribution

The novel approach integrates spectroscopic and photometric data using neural networks to improve stellar parameter estimation and dust reddening modeling.

Findings

Excellent fit to observed data across multiple bands.

Precise predictions of color-magnitude diagrams.

Improved, temperature-dependent reddening vector.

Abstract

We develop a data-driven model to map stellar parameters (effective temperature, surface gravity and metallicity) accurately and precisely to broad-band stellar photometry. This model must, and does, simultaneously constrain the passband-specific dust reddening vector in the Milky Way. The model uses a neural network to learn the (de-reddened) absolute magnitude in one band and colors across many bands, given stellar parameters from spectroscopic surveys and parallax constraints from Gaia. To demonstrate the effectiveness of this approach, we train our model on a dataset with spectroscopic parameters from LAMOST, APOGEE and GALAH, Gaia parallaxes, and optical and near-infrared photometry from Gaia, Pan-STARRS~1, 2MASS and WISE. Testing the model on these datasets leads to an excellent fit and a precise - and by construction accurate - prediction of the color-magnitude diagrams in many bands. This flexible approach rigorously links spectroscopic and photometric surveys, and also results in an improved, stellar-temperature-dependent reddening vector. As such, it provides a simple and accurate method for predicting photometry in stellar evolutionary models. Our model will form a basis to infer stellar properties, distances and dust extinction from photometric data, which should be of great use in 3D mapping of the Milky Way. Our trained model may be obtained at https://doi.org/10.5281/zenodo.3902382.