Stellar Parameters of BOSS M dwarfs in SDSS-V DR19

Basic Information

• Paper ID: 2511.20005
• Title: Stellar Parameters of BOSS M dwarfs in SDSS-V DR19
• Authors: Dan Qiu, Jennifer A. Johnson, Chao Liu, Diogo Souto, Ilija Medan and 18 other co-authors
• Classification: astro-ph.SR (Solar and Stellar Astrophysics), astro-ph.GA (Astrophysics of Galaxies)
• Publication Date: November 26, 2025 (Draft version)
• Paper Link: https://arxiv.org/abs/2511.20005v1

Abstract

This study derives stellar parameters (Fe/H, Teff, and log g) for M dwarf stars from low-resolution optical spectra (R~2000) obtained by the SDSS-V BOSS spectrograph using a data-driven model SLAM (Stellar LAbel Machine) based on Support Vector Regression (SVR). Metallicity Fe/H is calibrated using F/G/K dwarf companions from LAMOST, while effective temperature and surface gravity are calibrated using APOGEE Net. The study validates that SLAM predictions of Fe/H are unbiased in M+M binary systems with a scatter of 0.11 dex. Compared to other studies using F/G/K companion calibration, the biases are -0.06±0.16 dex and 0.02±0.14 dex, respectively. SLAM-derived effective temperatures are consistent with temperatures calibrated by interferometric angular diameters (bias of -27±92 K) and LAMOST temperatures (bias of -34±65 K), but are systematically lower than empirical relations based on color indices by 146±45 K. The study also provides equations to correct APOGEE ASPCAP metallicity.

Research Background and Motivation

Research Questions

M dwarf stars, as the most abundant stellar type in the Milky Way (comprising ~70% of all stars), require precise determination of stellar atmospheric parameters (Teff, log g, Fe/H) for understanding stellar physics, Galactic chemical evolution, and exoplanet research. However, the low-temperature atmospheres, complex molecular absorption features, and sensitivity to metallicity of M dwarfs present unique challenges for parameter determination.

Importance of the Problem

1. Galactic Archaeology: M dwarfs have long lifespans and their atmospheres retain chemical signatures of their formation environments, making them ideal tracers for reconstructing the Milky Way's chemical and dynamical history
2. Exoplanet Research: Many potentially habitable planets orbit M dwarf stars, and precise stellar parameters are crucial for planet characterization
3. Large Survey Requirements: SDSS-V has observed millions of M dwarf stars, necessitating reliable automated parameter determination methods

Limitations of Existing Methods

1. Insufficient Theoretical Models: Traditional 1D static model atmospheres (such as PHOENIX, BT-Settl) exhibit significant systematic errors due to:
   • Incomplete molecular and atomic line lists
   • Missing opacity sources
   • Failure of local thermodynamic equilibrium (LTE) assumptions in cool upper atmospheres
2. Optical Spectroscopy Challenges: Low-resolution optical regions are severely obscured by overlapping molecular features, making it difficult for theoretical models to decouple the effects of Teff, log g, and Fe/H
3. Lack of Calibration Samples: M dwarfs lack reference samples with precise metallicity measurements

Research Motivation

To bypass imperfect physical models using data-driven methods (SLAM) and establish reliable calibration samples through FGK+M wide binary systems (assuming both components share the same metallicity), thereby providing precise stellar parameters for the large-scale M dwarf sample in SDSS-V.

Core Contributions

1. Developed the SLAM Pipeline: Constructed a data-driven model based on SVR, integrated into the SDSS-V Astra analysis framework, capable of processing all BOSS M dwarf spectra
2. Established Calibration Sample: Identified and analyzed 1,120 FGK+M wide binary systems, utilizing F/G/K dwarfs from LAMOST for metallicity calibration and APOGEE Net for temperature and gravity calibration
3. Systematic Validation: Comprehensively validated SLAM parameter reliability through 256 M+M binary systems and cross-comparisons with multiple independent studies
4. Parameter Uncertainty Model: Established empirical relationships between parameter uncertainties and spectral signal-to-noise ratio (SNR), achieving ~0.19 dex (Fe/H), ~132 K (Teff), and ~0.1 dex (log g) at SNR=15
5. ASPCAP Correction Equations: Identified and quantified systematic deviations between SLAM and APOGEE ASPCAP metallicity, providing correction formulas dependent on Teff and Fe/H
6. Large-Scale Data Products: Provided stellar parameters for approximately 90,000 M dwarf candidates in SDSS-V DR19

Methodology Details

Task Definition

Input: BOSS low-resolution optical spectra (3800-9800 Å, R~2000)
Output: Stellar atmospheric parameters Fe/H, Teff, log g (and α/M)
Constraints: Parameter ranges limited to training set coverage: Fe/H ∈ -0.6, 0.5 dex, Teff ∈ 3100, 3900 K, log g ∈ 4.45, 4.95 dex

SLAM Model Architecture

1. Data Preprocessing

• Spectral Normalization: Using smooth splines (de Boor 1977) to fit the pseudo-continuum, dividing observed spectra by the pseudo-continuum
• Standardization: Normalizing both the normalized spectra and stellar labels to mean 0 and variance 1

2. Core Algorithm: Support Vector Regression (SVR)

SLAM employs radial basis functions (RBF) as the SVR kernel, training an independent SVR model for each spectral pixel.

Hyperparameter Optimization:

• C (penalty coefficient)
• ε (tube radius)
• γ (RBF kernel width)

Automatically optimized for each pixel by minimizing k-fold cross-validation mean squared error (CV MSE):

$\text{CV MSE}_j = \frac{1}{m}\sum_{i=1}^{m}[f_j(\vec{\theta}_i) - f_{i,j}]^2$

where $f_j(\vec{\theta}_i)$ is the model output for the i-th star at the j-th pixel, $f_{i,j}$ is the corresponding training spectrum value, and k=10.

3. Bayesian Parameter Inference

Utilizing Bayes' formula to maximize posterior probability:

$p(\vec{\theta}|\vec{f}_{\text{obs}}) \propto p(\vec{\theta})\prod_{j=1}^{n}p(f_{j,\text{obs}}|\vec{\theta})$

where $p(\vec{\theta})$ is the prior probability and $p(f_{j,\text{obs}}|\vec{\theta})$ is the likelihood function.

4. Uncertainty Estimation

Employing cross-validation scatter (CV scatter) as an empirical estimate of parameter uncertainty:

$\text{CV scatter} = \frac{1}{m}\sqrt{\sum_{i=1}^{m}(\vec{\theta}_{i,\text{SLAM}} - \vec{\theta}_i)^2}$

Establishing power-law relationships between uncertainty and SNR:

• $\sigma_{[\text{Fe/H}]} = 0.34 \times \text{SNR}^{-0.25}$
• $\sigma_{T_{\text{eff}}} = 388.88 \times \text{SNR}^{-0.47}$
• $\sigma_{\log g} = 0.13 \times \text{SNR}^{-0.13}$

Technical Innovations

1. Binary Star Calibration Strategy: Innovatively utilizing FGK+M wide binary systems, leveraging mature FGK parameter determination techniques to provide reliable metallicity benchmarks for M dwarfs
2. Mixed Label Sources:
   • Fe/H from LAMOST F/G/K companions (with temperature correction from Niu et al. 2023)
   • Teff and log g from APOGEE Net (III) near-infrared spectral analysis
   • Fully exploiting the advantages of different wavelength bands and instruments
3. Per-Pixel SVR: Compared to global fitting methods, per-pixel modeling better captures complex nonlinear relationships and local features
4. SNR-Dependent Uncertainties: Providing practical uncertainty estimation formulas applicable to SNR ranges of 2.3-19
5. Systematic Bias Correction: Addressing APOGEE ASPCAP's systematic metallicity underestimation, providing a two-dimensional correction function (dependent on Teff and Fe/H)

Experimental Setup

Datasets

M Dwarf Candidate Identification

Approximately 90,000 M dwarf candidates identified from SDSS-V DR19 using the following criteria:

1. BOSS spectral pipeline classification as M-type star
2. Combined CMD selection using Gaia photometry and color indices:
   • $M_{G0} > 6$ (excluding bright stars)
   • $M_{G0} < 3.75 \times (\text{BP\_RP}_0) + 7.50$ (excluding pre-main sequence and unresolved binaries)
   • $M_{G0} > 3.28 \times (\text{BP\_RP}_0) + 0.10$ (excluding white dwarfs)
   • ruwe < 1.4 (excluding unresolved binaries and poor astrometric data)

Training and Test Sets

• Training Set: 870 M dwarfs from FGK+M wide binary systems
  • Fe/H: LAMOST DR11 F/G/K companions (error <0.2 dex)
  • Teff, log g: APOGEE Net (III)
• Test Set: 250 M dwarfs (randomly divided)
• Validation Set: 256 M+M binary systems (for internal consistency checks)

Wide Binary Identification Criteria

Based on the Gaia eDR3 wide binary catalog from El-Badry et al. (2021):

• Projected separation < 1 pc
• Strict proper motion and parallax difference cutoffs
• Chance alignment probability $R_{\text{chance align}} < 0.1$
• Binary component ruwe < 1.4

Evaluation Metrics

1. Bias: Median difference, measuring systematic offset
2. Scatter: Standard deviation, measuring random error
3. Cross-Validation Metrics:
   • CV bias: Cross-validation bias
   • CV scatter: Cross-validation scatter

Comparison Methods

1. Metallicity:
   • Birky et al. (2020): The Cannon analysis of APOGEE spectra, calibrated with FGK companions
   • Behmard et al. (2025): The Cannon with FGK companion calibration
   • APOGEE ASPCAP DR19
2. Effective Temperature:
   • Birky et al. (2020): Optical spectra compared with BT-Settl models, calibrated with interferometric angular diameters
   • LAMOST gM/dM/sdM catalog (Du et al. 2024)
   • Mann et al. (2015, 2016): Empirical relations based on 2MASS and Gaia color indices
3. Surface Gravity:
   • LAMOST gM/dM/sdM catalog (Du et al. 2024)
   • Mann et al. (2016, 2019): Derived from stellar mass and radius

Implementation Details

• Spectral Resolution: R~2000 (BOSS)
• Wavelength Range: 3800-9800 Å
• Cross-Validation: k=10-fold
• SVR Kernel Function: Radial basis function (RBF)
• Hyperparameter Search: Grid search minimizing CV MSE
• Software Framework: Integrated into SDSS-V Astra v0.6.0

Experimental Results

Main Results

Test Set Performance

Performance on 250 test stars (Figure 4):

• Fe/H: Bias 0.03±0.25 dex
• Teff: Bias 11±168 K
• log g: Bias 0.00±0.10 dex

Results show no significant systematic trends, with residuals approximately symmetrically distributed and notably larger scatter in low-SNR spectra.

Metallicity Validation

1. M+M Binary Internal Consistency (Figure 6 left):
   • 256 M+M binary systems
   • Bias: -0.01±0.11 dex
   • Demonstrates SLAM's self-consistency in physically bound binary systems
2. Comparison with Birky et al. (2020) (Figure 6 middle):
   • Bias: -0.06±0.16 dex
   • Both use FGK companion calibration, showing excellent agreement
3. Comparison with Behmard et al. (2025) (Figure 6 right):
   • Bias: 0.02±0.14 dex
   • Excellent consistency, validating the calibration methodology

Effective Temperature Validation (Figure 7)

1. Comparison with Birky et al. (2020) (interferometric calibration):
   • Bias: -27±92 K
   • Excellent agreement
2. Comparison with LAMOST (Du et al. 2024):
   • Bias: -34±65 K
   • Very tight scatter of only 65 K
3. Comparison with Mann et al. (2015, 2016) (color-temperature relations):
   • Bias: -146±45 K
   • SLAM is systematically lower, possibly reflecting limitations of empirical relations

Surface Gravity Validation (Figure 8)

1. Comparison with LAMOST (Du et al. 2024):
   • Bias: -0.01±0.07 dex
   • Nearly perfect agreement
2. Comparison with Mass-Radius Derived (Mann et al. 2016, 2019):
   • Bias: -0.04±0.09 dex
   • Slight systematic trend at low log g (<4.7), but within acceptable range

Parameter Uncertainty Analysis (Figure 5)

Analyzing the relationship between CV scatter and SNR through equal-number SNR binning (~20 stars per bin):

• At SNR=15:
  • σFe/H ≈ 0.19 dex
  • σTeff ≈ 132 K
  • σlog g ≈ 0.1 dex
• Uncertainties decrease significantly with increasing SNR, consistent with expected statistical behavior

ASPCAP Metallicity Correction (Figures 10-11)

Systematic deviations between SLAM and APOGEE ASPCAP dependent on Teff and Fe/H are identified. A quadratic correction equation is provided:

$\Delta[\text{Fe/H}]_{\text{cal}} = -0.427x^2 - 0.158y^2 + 0.431xy + 2.767x - 2.156y - 4.405$

where $x = T_{\text{eff,ASPCAP}}/1000$ and $y = [\text{Fe/H}]_{\text{ASPCAP}}$

Comparison with PARSEC isochrones shows:

• ASPCAP original metallicity is systematically underestimated
• Corrected ASPCAP metallicity is consistent with PARSEC models and SLAM results
• Correction performs better at the metal-poor end (Fe/H<-0.6)

Experimental Findings

1. Importance of Calibration Method: The FGK companion calibration strategy is proven reliable, with three independent studies (this work, Birky, Behmard) using the same method obtaining consistent results
2. Complementarity of Wavelength Bands: Low-resolution optical spectra (BOSS) combined with high-resolution near-infrared spectral labels (APOGEE) effectively determine M dwarf parameters
3. Critical SNR Threshold: SNR>10 is a practical threshold for obtaining reliable parameters
4. Model Limitations: SLAM predictions have larger uncertainties in sparsely sampled training regions (Fe/H<-0.6, <20 training samples), and should not be extrapolated
5. ASPCAP Systematic Error: Confirms previously reported ASPCAP metallicity underestimation of approximately 0.10-0.24 dex (Qiu et al. 2024; Souto et al. 2022)

Related Work

M Dwarf Parameter Determination Methods

1. Theoretical Model Approaches:
   • PHOENIX, BT-Settl, MARCS and other 1D static model atmospheres
   • Limitations: Incomplete molecular line lists, LTE assumption failures, large systematic errors
2. High-Resolution Near-Infrared Spectroscopy:
   • APOGEE (Abdurro'uf et al. 2022)
   • CARMENES (Quirrenbach et al. 2016)
   • Advantages: Low molecular line density, clear continuum, reliable atomic lines
   • Limitations: Limited sample sizes
3. Data-Driven Methods:
   • The Payne/DD-Payne (Ting et al. 2019; Xiang et al. 2019)
   • The Cannon (Ness et al. 2015)
   • SLAM (Zhang et al. 2020; this work)
   • Advantages: Bypass imperfect physical models, suitable for large-scale data

Binary Star Calibration Studies

• Rojas-Ayala et al. (2010): Pioneering use of FGK+M binaries for calibration
• Mann et al. (2013): APOGEE high-resolution spectral analysis
• Montes et al. (2018): Optical spectral calibration
• Souto et al. (2020, 2022): APOGEE M dwarf metallicity benchmarks
• Qiu et al. (2024): LAMOST M dwarf parameter determination

SDSS-V MWM Project

• Objective: Resolve Galactic structure, composition, dynamics, and evolutionary history
• APOGEE: Near-infrared high-resolution (R~22,500), H<13
• BOSS: Optical low-resolution (R~2000), G≲20
• Astra Framework: Unified spectral analysis pipeline integrating multiple methods

Advantages of This Work

1. Sample Scale: Processes ~90,000 M dwarfs, far exceeding previous studies
2. Systematic Validation: Cross-validation with multiple independent datasets and methods
3. Practical Tools: Provides SNR-dependent uncertainty formulas and ASPCAP correction equations
4. Open Data: Parameters released in SDSS DR19, supporting community research

Conclusions and Discussion

Main Conclusions

1. SLAM Model Validity: Successfully developed and validated a data-driven SVR-based method capable of reliably determining M dwarf parameters from BOSS low-resolution optical spectra
2. Parameter Precision:
   • Fe/H: ~0.19 dex at SNR=15
   • Teff: ~132 K
   • log g: ~0.1 dex
3. Systematic Consistency: SLAM parameters are highly consistent with multiple independent studies, including:
   • M+M binary internal consistency (0.11 dex scatter)
   • Agreement with Birky and Behmard FGK companion calibration results
   • Alignment with LAMOST parameters
4. ASPCAP Correction: Identifies and quantifies ASPCAP metallicity systematic underestimation, providing practical correction formulas
5. Large-Scale Application: Provides reliable parameters for ~90,000 M dwarfs in SDSS-V DR19, integrated into the Astra analysis framework

Limitations

1. Parameter Range Restrictions:
   • SLAM cannot extrapolate, valid only within training domain
   • Valid ranges: Fe/H ∈ -0.6, 0.5 dex, Teff ∈ 3100, 3900 K, log g ∈ 4.45, 4.95 dex
   • Predictions outside training set (<20 stars) have large uncertainties
2. Metal-Poor Star Deficiency:
   • Sparse training samples for Fe/H<-0.6
   • Cannot determine whether this reflects survey selection effects or true astrophysical scarcity
   • Limits studies of Galactic halo M dwarfs
3. SNR Dependence:
   • Low-SNR (<10) spectra have significantly increased parameter uncertainties
   • Uncertainty formula extrapolation beyond SNR<2.3 and >19 may be inaccurate
4. Temperature Systematic Differences:
   • 146 K systematic difference with Mann et al. color-temperature relations
   • Root cause not fully clarified (possibly limitations of empirical relations or SLAM systematic errors)
5. Insufficient α/M Verification:
   • Although SLAM outputs α/M, insufficient external validation data exists

Future Directions

1. Expand Metal-Poor Training Set:
   • Search for additional FGK+M binaries in future SDSS releases (e.g., DR20)
   • Select metal-poor candidates through halo kinematics/orbital parameters
   • Conduct high-resolution optical/near-infrared follow-up observations for candidates lacking FGK companions
   • Specialized studies of metal-poor M dwarfs
2. Method Improvements:
   • Explore deep learning methods (e.g., neural networks) as SVR alternatives
   • Hybrid models combining physical models and data-driven approaches
   • Improved handling strategies for low-SNR spectra
3. Scientific Applications:
   • Galactic chemical evolution studies
   • Stellar population analysis
   • Age-metallicity-kinematics relations for M dwarfs
   • Exoplanet host star characterization
4. Multi-Wavelength Integration:
   • Joint analysis combining BOSS optical and APOGEE near-infrared spectra
   • Synergistic constraints utilizing Gaia photometry and astrometry
5. Uncertainty Quantification:
   • More refined uncertainty models considering local parameter space density
   • Complete posterior distributions within Bayesian framework

In-Depth Evaluation

Strengths

1. Methodological Innovation:
   • Clever exploitation of FGK+M binary systems to address M dwarf metallicity calibration challenges
   • Mixed label strategy fully leverages advantages of different instruments and wavelengths
   • Per-pixel SVR modeling captures complex spectral-parameter relationships
2. Experimental Sufficiency:
   • Multi-level validation: test set, M+M binaries, comparisons with multiple independent studies
   • Systematic analysis of SNR impact on uncertainties
   • Detailed ablation and sensitivity analyses
3. Result Convincingness:
   • Detailed numerical results with clear statistical significance
   • Consistency with multiple independent methods enhances credibility
   • Discovery and correction of ASPCAP systematic errors has important practical value
4. Practical Value:
   • Integration into SDSS official analysis pipeline (Astra)
   • Processing ~90,000 M dwarfs with publicly released data
   • Provides user-friendly uncertainty estimation formulas and ASPCAP correction equations
5. Writing Clarity:
   • Logical structure and clear reasoning
   • Detailed method descriptions ensuring reproducibility
   • Rich figures effectively conveying information

Shortcomings

1. Insufficient Physical Interpretation:
   • Lacks deep physical explanation for SLAM's effectiveness on low-resolution optical spectra
   • Insufficient analysis of temperature difference origins with Mann et al.
   • Unexplored spectral feature sensitivity to individual parameters
2. Metal-Poor Region Coverage:
   • Severely insufficient training samples for Fe/H<-0.6 (<20 stars)
   • Limits capability for Galactic halo and ancient stellar population studies
   • Cannot determine metal-poor M dwarf scarcity causes
3. Model Interpretability:
   • SVR as black-box model difficult to understand internal decision mechanisms
   • Unanalyzed wavelength regions' contribution to parameter determination
   • Missing feature importance analysis
4. Systematic Error Sources:
   • Label systematic errors (e.g., APOGEE Net uncertainties) insufficiently propagated
   • FGK companion metallicity temperature correction may introduce additional errors
   • Binary common origin assumptions may fail in extreme cases
5. Missing α/M Verification:
   • Although α/M parameters are output, paper acknowledges lacking validation data
   • Reliability of this important chemical abundance parameter questionable

Impact

1. Field Contributions:
   • Provides critical M dwarf parameter determination tool for SDSS-V
   • Establishes parameter catalog for ~90,000 M dwarfs, among the largest samples of this type
   • ASPCAP correction equations will benefit the community
2. Practical Value:
   • Integration into official analysis pipeline ensures long-term use
   • Open data and code promote reproducibility
   • Uncertainty formulas provide practical guidance for data users
3. Reproducibility:
   • Detailed method descriptions including hyperparameters and implementation details
   • Training data (FGK+M binary catalog) publicly released
   • Integration into Astra framework with accessible code
4. Potential Applications:
   • Galactic Archaeology: Tracing chemical evolution history
   • Exoplanet Science: Host star characterization
   • Stellar Physics: M dwarf atmospheric model validation
   • Dynamical Studies: 6D phase space analysis combined with Gaia data

Applicable Scenarios

1. Ideal Scenarios:
   • M dwarfs observed by SDSS-V BOSS (R~2000 optical spectra)
   • Main-sequence M dwarfs with parameters within training domain
   • Spectra with SNR>10
   • Studies requiring large-scale uniform parameter catalogs
2. Scenarios Requiring Caution:
   • Metal-poor stars (Fe/H<-0.6): Large uncertainties, potentially unreliable
   • Extreme parameters (Teff<3100 K or >3900 K): Extrapolation risks
   • Low-SNR (<5) spectra: Uncertainties may exceed nominal values
   • Non-main-sequence stars (e.g., pre-main sequence, subdwarfs): Unvalidated
3. Inapplicable Scenarios:
   • Studies requiring high-precision metallicity (σ<0.1 dex): High-resolution spectroscopy recommended
   • Studies requiring detailed chemical abundance patterns: SLAM provides only Fe/H and α/M
   • Non-BOSS spectra: Model trained on BOSS instrument characteristics

Methodological Insights

1. Data-Driven Astronomy: Demonstrates data-driven methods' powerful capability when theoretical models are insufficient
2. Multi-Source Label Fusion: Mixing labels from different sources (LAMOST metallicity, APOGEE Net temperature/gravity) is an effective strategy for complex problems
3. Binaries as Benchmarks: Physically bound binary systems are valuable resources for stellar parameter calibration
4. Cross-Validation Importance: Validation through multiple independent datasets is key to establishing credible results

References

Key Methodological Papers

1. Zhang et al. (2020): Original SLAM methodology paper
2. Ting et al. (2019): The Payne data-driven method
3. Ness et al. (2015): The Cannon method

Important Calibration Studies

4. Mann et al. (2013, 2015, 2016, 2019): M dwarf parameter determination series
5. Souto et al. (2020, 2022): APOGEE M dwarf metallicity benchmarks
6. Birky et al. (2020): APOGEE M dwarf parameters
7. Behmard et al. (2025): Latest M dwarf metallicity calibration

Surveys and Models

8. Kollmeier et al. (2025): SDSS-V overview
9. Allard et al. (2012): BT-Settl model atmospheres
10. El-Badry et al. (2021): Gaia wide binary catalog

Overall Assessment: This is a high-quality astronomical observational data analysis paper with reliable methods, sufficient validation, and high practical value. The SLAM model successfully addresses long-standing challenges in M dwarf parameter determination, providing critical tools for SDSS-V. The paper's main limitations are insufficient metal-poor region coverage and limited model interpretability, but clear improvement pathways exist for future work. This research will have significant impact on Galactic archaeology and exoplanet science.