LRPayne: Stellar parameters and abundances from low-resolution spectra

TL;DR

LRPayne introduces a model-driven neural-network framework to efficiently derive stellar parameters and chemical abundances from low-resolution optical spectra, addressing the data-volume challenge of surveys like WEAVE. It trains a 5-layer ANN on a library of 70,000 synthetic spectra generated with iSpec, MARCS atmospheres, and Turbospectrum, then uses chi-squared fitting to observed $R=5000$ spectra across 4200–6900 Å, with masking to mitigate synthetic gaps. Validation on Gaia FGK benchmark and metal-poor stars yields mean differences of $T_{ m eff}=22\pm87$ K, $\log g=0.19\pm0.23$ dex, and $[\mathrm{Fe/H}]=0.01\pm0.17$ dex, and reliably recovers abundances for Na, Mg, Si, Cr, Ni, V, and Sc among others; O, Mn in metal-rich giants, Al in metal-poor dwarfs, and log $g$ for hot metal-poor dwarfs remain challenging due to NLTE and line properties. The method achieves high interpolation accuracy on synthetic data ($<0.13\%$ for 90% of cases) and remains robust down to $S/N\sim30$, underscoring its potential for processing the WEAVE and similar survey data and enabling large-scale studies of Galactic chemical evolution.

Abstract

Aims. This paper introduces LRPayne, a novel algorithm designed for the efficient determination of stellar parameters and chemical abundances from low-resolution optical spectra, with a primary focus on data from large-scale galactic surveys such as WEAVE. Methods. LRPayne employs a model-driven approach, utilising a fully connected artificial neural network (ANN), trained on a library of 70,000 synthetic stellar spectra generated using iSpec with 1D MARCS model atmospheres and the Turbospectrum synthesis code. The network is trained to predict normalized flux given stellar labels (Teff, log(g), [Fe/H], vmic, vmax and v sin i, and 24 individual elemental abundances). Stellar parameters are subsequently derived from observed spectra by finding the best-fit synthetic spectrum from the ANN using a chi-squared minimisation technique. The method operates on spectra degraded to a resolution of R=5000 covering the wavelength range 4200-6900 Å. Results. Internal accuracy tests on synthetic spectra show a median interpolation error of less than 0.13 % for 90 % of the validation sample. The method accurately recovers most input labels from synthetic spectra, even at a signal-to-noise ratio (S/N) of 20, with some expected challenges for elements like Li, K, and N. Validation on observed spectra of 25 Gaia FGK benchmark stars and 42 metal-poor stars reveals good agreement with literature values. For stellar parameters, mean differences are 22+-87 K for Teff , 0.19+-0.23 dex for log(g), and 0.01+-0.17 dex for [Fe/H]. Abundances for elements like Na, Mg, Si, and most Fe-peak elements (Cr, Ni, V, Sc) are well-recovered. Challenges are noted for oxygen, manganese in metal-rich giants, aluminium in metal-poor stars and dwarfs, and for deriving log g in hot metal-poor dwarfs, partly due to non-local thermodynamic equilibrium effects and line characteristics.

Figures (8)

• Figure 1: Pictorial representation of the workflow of LRPayne (left) and architecture of the artificial neural network used with LRPayne (right).
• Figure 2: Solar spectrum (black) with masked pixels show with red vertical lines. Important masked lines and data gap are shown in text.
• Figure 3: Left: Histogram representing median interpolation error of the ANN for 5000 synthetic models; Middle: Co-relation of interpolation error on effective temperature and surface gravity of the synthetic models; Right: Co-relation of interpolation error on metallicity and surface gravity of the synthetic models.
• Figure 4: Comparison of input labels with respect to the values recovered by LRPayne though fitting multiple synthetic spectra at different S/N (Green : S/N = 20, Red : S/N = 100 and Black : S/N = 1000). Note: Only 14 elements [O, Na, Al, Mg, Si, Ti, Ca, Mn, Cr, Ni, V, Sc, Ba and Y] have been validated using observed data (See Section \ref{['Sec:abun']}). For other elements, this result represents the upper limit on the sensitivity of LRPayne
• Figure 5: Parameters inferred for Gaia FGK benchmark (red circle) and metal-poor stars (blue plus sign) using LRPayne in comparison with the literature. The solid black line represents zero difference between LRPayne derived value and literature, whereas the two dashed black line represents a range within which the accuracy of LRPayne is comparable to traditional methods of analysis. Literature parameters for benchmark stars were obtained from jofre2014 and heiter2015, and for metal-poor stars from bensby2014. The mean and 1$\sigma$ of the distribution is given on the top left of each panel.