A New Hybrid Machine Learning Method for Stellar Parameter Inference

Abstract The advent of machine learning (ML) is revolutionary to numerous scientific disciplines, with a growing number of examples in astronomical spectroscopic inference, as ML is more powerful than traditional techniques. Here we introduce a hybrid ML (HML) method combining automatic differentiation, interpolation, and Bayesian optimization to infer stellar parameters of stellar spectra. We study T eff , <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mi>log</mml:mi> <mml:mo stretchy="false">(</mml:mo> <mml:mi>g</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:math> , and [Fe/H], but this method could be extended to other parameters such as [ α /Fe] (alpha element abundance), C/O (carbon–oxygen ratio), and f sed (sedimentation efficiency). We first use blase to semiempirically recast spectra into sets of Voigt profiles. blase is run on 1314 synthetic spectra from a rectilinear subset of the PHOENIX model grid ( T eff : [2300, 12,000] K, <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mi>log</mml:mi> <mml:mo stretchy="false">(</mml:mo> <mml:mi>g</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:math> : [2, 6], [Fe/H]: [–0.5, 0], λ : [8038, 12849] Å). For 128,723 detected features, we map stellar parameters to spectral line parameters using linear interpolation. This creates the PHOENIX generator, enabling parallelized spectral synthesis. Gaussian process minimization is used to infer stellar parameters by minimizing a rms loss function. Testing 210 noise-free models ( T eff : [3000, 11,000] K, <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mi>log</mml:mi> <mml:mo stretchy="false">(</mml:mo> <mml:mi>g</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:math> : [2, 6], [Fe/H]: [−0.5, 0]), we find inference errors: T eff : 93 K, <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mi>log</mml:mi> <mml:mo stretchy="false">(</mml:mo> <mml:mi>g</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:math> : 0.24, and [Fe/H]: 0.056 for T eff &lt; 7000 K, and T eff : 347 K, <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" overflow="scroll"> <mml:mi>log</mml:mi> <mml:mo stretchy="false">(</mml:mo> <mml:mi>g</mml:mi> <mml:mo stretchy="false">)</mml:mo> </mml:math> : 0.26, and [Fe/H]: 0.16 for T eff ≥ 7000 K. We also upload online an archive of blase models of the PHOENIX subset. This proof-of-concept study shows that semiempirical HML is a viable alternative to traditional approaches in spectroscopic inference.