Using autoencoders and deep transfer learning to determine the stellar parameters of 286 CARMENES M dwarfs

TL;DR

The paper tackles the challenge of determining fundamental parameters ($T_{\rm eff}$, $\\log g$, [Fe/H], $v\\sin{i}$) for M-dwarfs from high-resolution spectra in the presence of a synthetic gap between synthetic PHOENIX-ACES spectra and observed CARMENES data. It introduces a feature-based deep transfer learning framework using autoencoders to map both synthetic and observed spectra into a shared latent space, followed by CNN regression to estimate stellar parameters for 286 stars. The method reduces domain differences in latent space and yields parameter estimates broadly consistent with recent CARMENES studies, with improved metallicity behavior due to the DTL approach and a Teff tilt at higher temperatures. It explores astrophysical implications with activity and kinematic population diagnostics and provides open-source code for the methodology. The work demonstrates the potential of DTL to bridge domain gaps in stellar spectroscopy and provides open-source code.

Abstract

Deep learning (DL) techniques are a promising approach among the set of methods used in the ever-challenging determination of stellar parameters in M dwarfs. In this context, transfer learning could play an important role in mitigating uncertainties in the results due to the synthetic gap (i.e. difference in feature distributions between observed and synthetic data). We propose a feature-based deep transfer learning (DTL) approach based on autoencoders to determine stellar parameters from high-resolution spectra. Using this methodology, we provide new estimations for the effective temperature, surface gravity, metallicity, and projected rotational velocity for 286 M dwarfs observed by the CARMENES survey. Using autoencoder architectures, we projected synthetic PHOENIX-ACES spectra and observed CARMENES spectra onto a new feature space of lower dimensionality in which the differences between the two domains are reduced. We used this low-dimensional new feature space as input for a convolutional neural network to obtain the stellar parameter determinations. We performed an extensive analysis of our estimated stellar parameters, ranging from 3050 to 4300 K, 4.7 to 5.1 dex, and -0.53 to 0.25 dex for Teff, logg, and [Fe/H], respectively. Our results are broadly consistent with those of recent studies using CARMENES data, with a systematic deviation in our Teff scale towards hotter values for estimations above 3750 K. Furthermore, our methodology mitigates the deviations in metallicity found in previous DL techniques due to the synthetic gap. We consolidated a DTL-based methodology to determine stellar parameters in M dwarfs from synthetic spectra, with no need for high-quality measurements involved in the knowledge transfer. These results suggest the great potential of DTL to mitigate the differences in feature distributions between the observations and the PHOENIX-ACES spectra.

Figures (23)

• Figure 1: Two-dimensional UMAP projection of PHOENIX-ACES (dots colour-coded by $T_{\rm eff}$) and CARMENES (grey triangles) spectra from the 8 800--8 835 $\AA$ window. Almost all CARMENES spectra are isolated from the PHOENIX-ACES family feature space.
• Figure 2: Schematic representation of the AE architecture used in this work.
• Figure 3: Reconstructed spectrum (left) and latent representation (right) of a PHOENIX-ACES synthetic spectrum for one of the trained AEs. Left panel: comparison of the original (blue) and reconstructed (red) spectrum. Both spectra overlap as they are almost similar. The title shows the stellar parameters of the synthetic spectrum. Reconstruction residuals (original$-$reconstructed) are shown in the bottom panel. Right panel: 32-dimensional latent space of the input spectrum obtained by the encoder, reshaped to a 8$\times$4 matrix only for a better visibility. The colour scale indicates the strength of the features. The decoder uses this compressed representation to obtain the reconstructed spectrum.
• Figure 4: Original (blue) vs. reconstructed CARMENES spectrum for LSPM J0422+1031 (Karmn J04225+105, M3.5 V). The Figure only shows a section of the spectrum for better visibility, with the unique purpose of emphasising how the reconstruction after fine-tuning (black) captures much more detailed spectral features than the reconstruction with the initial training (red).
• Figure 5: Two-dimensional UMAP projection of one of the 26 sets of PHOENIX-ACES (dots colour-coded by $T_{\rm eff}$) and CARMENES (grey triangles) compressed representations. PHOENIX-ACES encodings are obtained with the initially trained AE and CARMENES encodings with the fine-tuned network.