Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Estimating stellar atmospheric parameters and elemental abundances using fully connected residual network

Shuo Li, Yin-Bi Li, A-Li Luo, Jun-Chao Liang, Hai-Ling Lu, Hugh R. A. Jones

TL;DR

This work tackles the challenge of inferring stellar atmospheric parameters and elemental abundances from ultra-low-resolution spectra ($R\sim200$) by training a Fully Connected Residual Network (FCResNet) on LAMOST spectra degraded to CSST-like resolution with APOGEE labels. FCResNet combines a simple fully connected architecture with residual connections to enable stable training and high predictive accuracy, outperforming KNN, XGBoost, SVR, and CNN in predicting $T_ ext{eff}$, $\log g$, [Fe/H], [C/Fe], [N/Fe], and [Mg/Fe], while remaining computationally efficient (42 s for 10^6 spectra) and lightweight (348 KB). The model demonstrates robust performance across varying $S/N_g$, strong agreement with APOGEE and PARSEC isochrones in Kiel diagrams, and interpretable spectral feature usage via SHAP (e.g., MgI b at ~5184 Å). The results indicate FCResNet is a practical, scalable tool for CSST's large-volume ultra-low-resolution data, with potential to extend to broader abundance analyses and guiding future survey pipelines.

Abstract

Stellar atmospheric parameters and elemental abundances are traditionally determined using template matching techniques based on high-resolution spectra. However, these methods are sensitive to noise and unsuitable for ultra-low-resolution data. Given that the Chinese Space Station Telescope (CSST) will acquire large volumes of ultra-low-resolution spectra, developing effective methods for ultra-low-resolution spectral analysis is crucial. In this work, we investigated the Fully Connected Residual Network (FCResNet) for simultaneously estimating atmospheric parameters ($T_\text{eff}$, $\log g$, [Fe/H]) and elemental abundances ([C/Fe], [N/Fe], [Mg/Fe]). We trained and evaluated FCResNet using CSST-like spectra (\textit{R} $\sim$ 200) generated by degrading LAMOST spectra (\textit{R} $\sim$ 1,800), with reference labels from APOGEE. FCResNet significantly outperforms traditional machine learning methods (KNN, XGBoost, SVR) and CNN in prediction precision. For spectra with g-band signal-to-noise ratio greater than 20, FCResNet achieves precisions of 78 K, 0.15 dex, 0.08 dex, 0.05 dex, 0.10 dex, and 0.05 dex for $T_\text{eff}$, $\log g$, [Fe/H], [C/Fe], [N/Fe] and [Mg/Fe], respectively, on the test set. FCResNet processes one million spectra in only 42 seconds while maintaining a simple architecture with just 348 KB model size. These results suggest that FCResNet is a practical and promising tool for processing the large volume of ultra-low-resolution spectra that will be obtained by CSST in the future.

Estimating stellar atmospheric parameters and elemental abundances using fully connected residual network

TL;DR

This work tackles the challenge of inferring stellar atmospheric parameters and elemental abundances from ultra-low-resolution spectra () by training a Fully Connected Residual Network (FCResNet) on LAMOST spectra degraded to CSST-like resolution with APOGEE labels. FCResNet combines a simple fully connected architecture with residual connections to enable stable training and high predictive accuracy, outperforming KNN, XGBoost, SVR, and CNN in predicting , , [Fe/H], [C/Fe], [N/Fe], and [Mg/Fe], while remaining computationally efficient (42 s for 10^6 spectra) and lightweight (348 KB). The model demonstrates robust performance across varying , strong agreement with APOGEE and PARSEC isochrones in Kiel diagrams, and interpretable spectral feature usage via SHAP (e.g., MgI b at ~5184 Å). The results indicate FCResNet is a practical, scalable tool for CSST's large-volume ultra-low-resolution data, with potential to extend to broader abundance analyses and guiding future survey pipelines.

Abstract

Stellar atmospheric parameters and elemental abundances are traditionally determined using template matching techniques based on high-resolution spectra. However, these methods are sensitive to noise and unsuitable for ultra-low-resolution data. Given that the Chinese Space Station Telescope (CSST) will acquire large volumes of ultra-low-resolution spectra, developing effective methods for ultra-low-resolution spectral analysis is crucial. In this work, we investigated the Fully Connected Residual Network (FCResNet) for simultaneously estimating atmospheric parameters (, , [Fe/H]) and elemental abundances ([C/Fe], [N/Fe], [Mg/Fe]). We trained and evaluated FCResNet using CSST-like spectra (\textit{R} 200) generated by degrading LAMOST spectra (\textit{R} 1,800), with reference labels from APOGEE. FCResNet significantly outperforms traditional machine learning methods (KNN, XGBoost, SVR) and CNN in prediction precision. For spectra with g-band signal-to-noise ratio greater than 20, FCResNet achieves precisions of 78 K, 0.15 dex, 0.08 dex, 0.05 dex, 0.10 dex, and 0.05 dex for , , [Fe/H], [C/Fe], [N/Fe] and [Mg/Fe], respectively, on the test set. FCResNet processes one million spectra in only 42 seconds while maintaining a simple architecture with just 348 KB model size. These results suggest that FCResNet is a practical and promising tool for processing the large volume of ultra-low-resolution spectra that will be obtained by CSST in the future.

Paper Structure

This paper contains 11 sections, 3 equations, 10 figures, 3 tables.

Table of Contents

  1. Introduction
  2. Data
  3. Method
  4. Fully Connected Neural Network
  5. Residual Blocks
  6. FCResNet
  7. Performance Evaluation
  8. Evaluation Criteria
  9. Results
  10. Comparison of different algorithms
  11. Summary

Figures (10)

  • Figure 1: Examples of abnormal LAMOST spectra. The top panel shows a spectrum with missing flux, while the bottom panel shows a spectrum with abnormal flux.
  • Figure 2: Comparison between the original LAMOST spectrum (black), covering the range 3690 – 9100 Å, and the CSST-like spectrum (blue), covering 4000 – 8122 Å. The CSST-like spectrum was obtained by degrading the original resolution from R = 1800 to R = 200, following the method described in 2019JOSS....4.1387L.
  • Figure 3: Distributions of stellar parameters in our dataset. From left to right and top to bottom, the panels correspond to $T_\text{eff}$, $\log g$, [Fe/H], [C/Fe], [N/Fe] and [Mg/Fe], respectively. These parameters were derived from APOGEE DR16 and correspond to 22,632 stars cross-matched with LAMOST DR8. All parameters were selected based on rigorous quality criteria to ensure the reliability of the sample for model training and evaluation.
  • Figure 4: Left: The structure of the residual block used in this work. Right: The architecture of the proposed Fully Connected Residual Network (FCResNet) model.
  • Figure 5: Comparison between FCResNet predictions and APOGEE labels on the test set. From left to right and top to bottom, the panels correspond to $T_\text{eff}$, $\log g$, [Fe/H], [C/Fe], [N/Fe], [Mg/Fe], [C/H], [N/H] and [Mg/H], respectively. The first six parameters are directly predicted by the model, while the last three are calculated using [X/H] = [X/Fe] + [Fe/H]. In each panel, the upper subplot shows a density scatter plot of FCResNet predictions versus APOGEE labels, while the lower subplot displays the residuals (predicted value minus label) versus the labels. The $\mu$ and $\sigma$ represent the systematic bias and dispersion, respectively.
  • ...and 5 more figures