TheUse of Conditional Variational Autoencoders in Generating Stellar Spectra

TL;DR

This paper introduces a conditional variational autoencoder (CVAE) that rapidly generates accurate stellar spectra across a wide range of stellar parameters, serving as an efficient surrogate for traditional radiative transfer models.

Contribution

The authors develop and validate a CVAE that synthesizes stellar spectra faster than line-by-line methods with high accuracy, enabling real-time spectral modeling.

Findings

Median absolute residual <$1.8×10^{-3}$ flux units

No wavelength-dependent bias in residuals

Effective across diverse stellar parameters

Abstract

We present a conditional variational autoencoder (CVAE) that generates stellar spectra covering 4000 $\le$ $T_{\mathrm{eff}$ $\le$ 11,000 K, $2.0 \le \log g \le 5.0$ dex, $-1.5 \le [\mathrm{M}/\mathrm{H}] \le +1.5$ dex, $v\sin i \le 300$ km/s, $\xi_t$ between 0 and 4 km/s, and for any instrumental resolving powers less than 115,000. The spectra can be calculated in the wavelength range 4450-5400 \AA. Trained on a grid of \textsc{SYNSPEC} spectra, the network synthesizes a spectrum in around two orders of magnitude faster than line-by-line radiative transfer. We validate the CVAE on $10^4$ test spectra unseen during training. Pixel-wise statistics yield a median absolute residual of <$1.8\times10^{-3}$ flux units with no wavelength-dependent bias. A residual error map across the parameters plane shows $\langle|\Delta F|\rangle<2\times10^{-3}$ everywhere, and marginal diagnostics versus $T_{\mathrm{eff}}$, $\log g$, $v\sin i$, $\xi_t$, and $[Fe/H]$\ reveal no relevant trends. These results demonstrate that the CVAE can serve as a drop-in, physics-aware surrogate for radiative transfer codes, enabling real-time forward modeling in stellar parameter inference and offering promising tools for spectra synthesis for large astrophysical data analysis.