[Paper Review] Metallicities in M dwarfs: Investigating different determination techniques

V. M. Passegger, A. Bello-García+2 more|arXiv (Cornell University)|Feb 24, 2022

Stellar, planetary, and galactic studies91 references46 citations

TL;DR

This study evaluates four methods—synthetic spectral fitting, pseudo-equivalent width analysis, and machine learning—for determining stellar parameters in M dwarfs using CARMENES high-resolution spectra. Despite standardizing key components like model spectra and wavelength ranges, significant discrepancies in effective temperature (up to 200 K) and metallicity (up to 0.3 dex) persist, indicating that standardization alone is insufficient for consistency in M dwarfs, highlighting the need for deeper methodological refinement beyond surface-level calibration.

ABSTRACT

Deriving metallicities for solar-like stars follows well-established methods, but for cooler stars such as M dwarfs, the determination is much more complicated due to forests of molecular lines that are present. Several methods have been developed in recent years to determine accurate stellar parameters for these cool stars (T-eff less than or similar to 4000 K). However, significant differences can be found at times when comparing metallicities for the same star derived using different methods. In this work, we determine the effective temperatures, surface gravities, and metallicities of 18 well-studied M dwarfs observed with the CARMENES high-resolution spectrograph following different approaches, including synthetic spectral fitting, analysis of pseudo-equivalent widths, and machine learning. We analyzed the discrepancies in the derived stellar parameters, including metallicity, in several analysis runs. Our goal is to minimize these discrepancies and find stellar parameters that are more consistent with the literature values. We attempted to achieve this consistency by standardizing the most commonly used components, such as wavelength ranges, synthetic model spectra, continuum normalization methods, and stellar parameters. We conclude that although such modifications work quite well for hotter main-sequence stars, they do not improve the consistency in stellar parameters for M dwarfs, leading to mean deviations of around 50-200 K in temperature and 0.1-0.3 dex in metallicity. In particular, M dwarfs are much more complex and a standardization of the aforementioned components cannot be considered as a straightforward recipe for bringing consistency to the derived parameters. Further in-depth investigations of the employed methods would be necessary in order to identify and correct for the discrepancies that remain.

Motivation & Objective

• To assess the consistency of stellar parameter determinations—especially metallicity—across multiple independent methods for M dwarfs.
• To investigate whether standardizing key components (wavelength ranges, model spectra, continuum normalization, and stellar parameters) reduces discrepancies between methods.
• To identify the root causes of persistent inconsistencies in derived parameters despite standardization.
• To evaluate the performance of machine learning and spectral synthesis tools in achieving consistency with literature values.
• To determine whether method-specific improvements can reduce discrepancies and enhance reliability in M dwarf parameter determination.

Proposed method

• Applied four distinct methods: synthetic spectral fitting (SteParSyn, ODUSSEAS, Pass19-code), pseudo-equivalent width (pEW) measurements, and deep learning (DL) regression on high-resolution CARMENES spectra.
• Conducted four analysis runs: Run A (unrestricted method use), Run B (fixed T_eff and log g to derive only [Fe/H]), and Runs C/C2 (standardized model spectra, continuum normalization, and wavelength ranges).
• Used synthetic model atmospheres (BT-Settl and PHOENIX-ACES) and radiative transfer codes (turbospectrum, SME) to generate synthetic spectra for fitting.
• Employed χ²-minimization and machine learning models trained on spectral features to derive T_eff, log g, and [Fe/H] across different wavelength ranges.
• Compared results across methods and runs against literature median values and assessed consistency via ΔT_eff and Δ[Fe/H] metrics.
• Evaluated method-specific improvements, such as extending wavelength coverage for DL and adjusting line sets for ODUSSEAS.

Experimental results

Research questions

• RQ1 To what extent does standardizing model spectra, wavelength ranges, and continuum normalization reduce discrepancies in T_eff and [Fe/H] among different analysis methods for M dwarfs?
• RQ2 Why do significant discrepancies (up to 200 K in T_eff and 0.3 dex in [Fe/H]) persist even after standardization of key components?
• RQ3 How do the performances of machine learning and traditional spectral synthesis methods compare in terms of consistency with literature values and among themselves?
• RQ4 Which method-specific adjustments (e.g., extended wavelength ranges, new reference T_eff scales) can most effectively improve parameter consistency?
• RQ5 To what extent do differences in stellar atmosphere models and line lists contribute to the observed parameter discrepancies?

Key findings

• Standardization of model spectra, wavelength ranges, and continuum normalization did not improve consistency in stellar parameters for M dwarfs, with mean deviations reaching 50–200 K in T_eff and 0.1–0.3 dex in [Fe/H] in Runs C and C2.
• The original, unrestricted method settings (Run A) showed the best agreement with literature median values, with T_eff differences below 100 K and [Fe/H] differences below 0.1 dex for Pass19-code and SteParSyn.
• Machine learning (DL) methods produced systematically higher [Fe/H] values and would benefit from using multiple wavelength ranges to improve consistency.
• ODUSSEAS consistently derived lower T_eff values, and its performance could be enhanced by adopting a new reference T_eff scale based on interferometry.
• Discrepancies persist due to methodological differences beyond standardization, particularly in stellar atmosphere models, line lists, and equations of state, which require deeper investigation.
• No universal recipe exists for achieving consistency; method-specific refinements and deeper scrutiny of underlying assumptions are essential to reduce parameter discrepancies in M dwarfs.