Generalization from Low- to Moderate-Resolution Spectra with Neural Networks for Stellar Parameter Estimation: A Case Study with DESI

Xiaosheng Zhao; Yuan-Sen Ting; Rosemary F. G. Wyse; Alexander S. Szalay; Yang Huang; László Dobos; Tamás Budavári; Viska Wei

astro-ph.SRFeb 16, 2026

Generalization from Low- to Moderate-Resolution Spectra with Neural Networks for Stellar Parameter Estimation: A Case Study with DESI

Authors: Xiaosheng Zhao, Yuan-Sen Ting, Rosemary F. G. Wyse, Alexander S. Szalay, Yang Huang, László Dobos, Tamás Budavári, Viska Wei

Organizations: Department of Physics & Astronomy, The Johns Hopkins University, Baltimore, MD 21218, USA · Department of Astronomy, The Ohio State University, 140 West 18th Avenue, Columbus, OH 43210, USA · Center for Cosmology and AstroParticle Physics (CCAPP), The Ohio State University, Columbus, OH 43210, USA · Max-Planck-Institut für Astronomie, Königstuhl 17, D-69117 Heidelberg, Germany · Department of Computer Science, The Johns Hopkins University, Baltimore, MD 21218, USA · School of Astronomy and Space Science, University of Chinese Academy of Sciences, Beijing 100049, People’s Republic of China · National Astronomical Observatories, Chinese Academy of Sciences, Beijing 100012, People’s Republic of China · Department of Information Systems, Eötvös Loránd University, Budapest 1117, Hungary · Department of Applied Mathematics & Statistics, Johns Hopkins University, Baltimore, MD 21218, USA

Abstract

Cross-survey generalization is a critical challenge in stellar spectral analysis, particularly in cases such as transferring from low- to moderate-resolution surveys. We investigate this problem using pre-trained models, focusing on simple neural networks such as multilayer perceptrons (MLPs), with a case study transferring from LAMOST low-resolution spectra (LRS) to DESI medium-resolution spectra (MRS). Specifically, we pre-train MLPs on either LRS or their embeddings and fine-tune them for application to DESI stellar spectra. We compare MLPs trained directly on spectra with those trained on embeddings derived from transformer-based models (self-supervised foundation models pre-trained for multiple downstream tasks). We also evaluate different fine-tuning strategies, including residual-head fine-tuning, LoRA, and full fine-tuning. We find that MLPs pre-trained on LAMOST LRS achieve strong performance, even without fine-tuning, and that modest fine-tuning with DESI spectra further improves the results. For iron abundance, embeddings from a transformer-based model yield advantages in the metal-rich ([Fe/H] > -1.0) regime, but underperform in the metal-poor regime compared to MLPs trained directly on LRS. We also show that the optimal fine-tuning strategy depends on the specific stellar parameter under consideration. These results highlight that simple pre-trained MLPs can provide competitive cross-survey generalization, while the role of spectral foundation models for cross-survey stellar parameter estimation requires further exploration.

Figures & tables

Figure 1: Sketch of the pre-training and fine-tuning workflow. Pre-training uses normalized spectra or foundation-model spectral embeddings to predict labels: [Fe/H] from APOGEE ( $>$ –2.0), supplemented at lower metallicities by PASTEL, SAGA, and other VMP/UMP datasets; [ $\alpha$ /Fe] from APOGEE. The foundation model is trained on large unlabeled datasets. Fine-tuning then adapts the network using fewer labeled spectra, either from spectra or spectral embeddings, with different fine-tuning modules applied (see Section III ).

Figure 2: Comparison of [Fe/H] and [ $\alpha$ /Fe] estimates between the DESI SP pipeline and MLP-based models, referenced against APOGEE DR17 labels. From left to right: DESI SP –DESI SP pipeline, MLP-Scratch –MLP trained from scratch on the same number of DESI spectra used for fine-tuning, MLP-LRS –MLP pre-trained on LAMOST LRS (zero-shot application), and MLP-LRS (fine-tuned) –MLP pre-trained on LAMOST LRS and fine-tuned with DESI spectra. The first row shows [Fe/H] results; the second row shows [ $\alpha$ /Fe]. Legends list the number of test stars (N), coefficient of determination ( $\mathrm{R^{2}}$ ), and the robustly estimated standard deviation of the residuals ( $\sigma$ , computed after 3 $\sigma$ clipping with sigma_clip in astropy ). All sources are restricted to $T_{\rm eff}>4000$ K. The dashed black line indicates the ideal 1:1 relation, and the gray dotted lines mark ±0.2 dex deviations. The fine-tuning set consists of 2,069 stars. A comparison using the clean, calibrated DESI SP subset is provided in Appendix D .

Figure 3: [ $\alpha$ /Fe]–[Fe/H] diagrams for DESI EDR and APOGEE reference sample. From left to right: (1) DESI SP : DESI SP pipeline, (2) MLP-Scratch : MLP trained from scratch on the same number of DESI spectra used for fine-tuning, (3) MLP-LRS : MLP pre-trained on LAMOST LRS (zero-shot application), (4) MLP-LRS (fine-tuned) : MLP pre-trained on LAMOST LRS and fine-tuned with DESI spectra, and (5) APOGEE : DESI DR1–APOGEE DR17 cross-matched sample (7,608 stars). The first row shows raw counts, and the second row shows column-normalized densities, where each [Fe/H] column is divided by its maximum count. The legend reports the Maximum Mean Discrepancy (MMD) between each sample and the APOGEE reference.

Figure 4: Effect of fine-tuning sample size (0, 100, 200, 500, 1000, and 2069, where 0 denotes zero-shot performance), evaluated on the testing sample. Shown are $R^{2}$ and the robustly estimated standard deviation of the residuals ( $\sigma$ ) for [Fe/H] and [ $\alpha$ /Fe], obtained after fine-tuning (residual-head fine-tuning for [Fe/H] and LoRA fine-tuning for [ $\alpha$ /Fe]). For both metrics, we plot the mean and 1 $\sigma$ error bars from five independent runs. The scatter plots for the different sample sizes are shown in Figure A1 and A2 .

[Fe/H]
(Full/MP/MR)	lrs	embed	embed-align	embed-split
residual	0.923 / 0.636 / 0.904	0.922 / 0.196 / 0.913	0.922 / 0.003 / 0.918	0.933 / 0.219 / 0.927
LoRA	0.922 / 0.571 / 0.904	0.921 / 0.031 / 0.916	0.928 / 0.131 / 0.922	0.933 / 0.133 / 0.930
full fine-tune	0.923 / 0.329 / 0.911	0.926 / 0.173 / 0.918	0.927 / 0.085 / 0.922	0.933 / 0.079 / 0.931
zero-shot	0.873 / 0.678 / 0.833	0.833 / 0.694 / 0.779	0.851 / 0.465 / 0.810	0.839 / 0.537 / 0.791
from scratch	0.880 / -0.736 / 0.878	0.877 / -1.246 / 0.886	0.647 / -0.072 / 0.542	0.794 / -0.649 / 0.759

Table 1: Coefficient of determination (

R^{2}

) for MLPs with different fine-tuning strategies and input types in [Fe/H] and [

\alpha

/Fe] estimation on DESI spectra.

Figure 5: Loss landscapes for different fine-tuning strategies, evaluated on the metal-poor regime of [Fe/H] (79 stars). The first row shows 3D surfaces, and the second row shows 2D contours of the logarithmic MSE loss, plotted as a function of perturbations around the pre-trained model parameters along two random directions. The minimum losses are indicated in the colorbar of the 2D contours.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A1: Effect of fine-tuning sample size on the one-to-one comparisons, evaluated on all remaining samples not used for fine-tuning (See Table A1 for comparisons using a fixed testing sample). Results are shown for residual-head fine-tuning of [Fe/H] and LoRA fine-tuning of [ $\alpha$ /Fe]. Other conventions follow Figure 2 .

Figure A2: Effect of fine-tuning sample size on the predicted [ $\alpha$ /Fe]–[Fe/H] diagrams, evaluated on all remaining samples not used for fine-tuning. Results are shown for residual-head fine-tuning of [Fe/H] and LoRA fine-tuning of [ $\alpha$ /Fe]. Other conventions follow Figure 3 .

[Fe/H]
M-poor:M-rich	243:1826	117:883	59:441
lrs	0.923 / 0.636 / 0.904	0.916 / 0.619 / 0.894	0.910 / 0.600 / 0.886
embed	0.922 / 0.196 / 0.913	0.920 / 0.160 / 0.911	0.913 / 0.216 / 0.900
embed-align	0.922 / 0.003 / 0.918	0.922 / -0.026 / 0.918	0.910 / -0.233 / 0.907
embed-split	0.933 / 0.219 / 0.927	0.927 / 0.058 / 0.924	0.926 / 0.213 / 0.917
from scratch	0.880 / -0.736 / 0.878	0.826 / -1.663 / 0.826	0.433 / -3.575 / 0.336

Appendix

Table A1: Coefficient of determination (

R^{2}

) for MLPs with different input types as a function of fine-tuning sample size.

[Fe/H]
M-poor:M-rich	243:1826	117:883	59:441
residual	0.923 / 0.636 / 0.904	0.916 / 0.619 / 0.894	0.910 / 0.600 / 0.886
LoRA	0.922 / 0.571 / 0.904	0.912 / 0.496 / 0.891	0.914 / 0.590 / 0.892
full fine-tune	0.923 / 0.329 / 0.911	0.916 / 0.401 / 0.900	0.915 / 0.516 / 0.895
scratch	0.880 / -0.736 / 0.878	0.826 / -1.663 / 0.826	0.433 / -3.575 / 0.336
M-poor:M-rich	23:177	12:88	0:0

Appendix

Table A2: Coefficient of determination (

R^{2}

) for MLPs with different fine-tuning strategies as a function of fine-tuning sample size.

Figure A3: Effect of the number of trainable parameters on [Fe/H] fine-tuning performance, comparing LoRA and residual-head strategies. The y-axis shows the coefficient of determination ( $R^{2}$ ) for [Fe/H] estimation using the full sample, metal-rich subset ([Fe/H] $>-1.0$ ), and metal-poor subset ([Fe/H] $<-1.0$ ). The number of test samples is 5,539, with a fine-tuning set of 2,069 stars. Annotations indicate the hidden-layer size for residual-head models (e.g., H384 ) and the rank for LoRA models (e.g., R128 ).

Figure A4: Effect of learning rate (LR) on models trained from scratch for [Fe/H] and [ $\alpha$ /Fe], referenced against APOGEE. Conventions follow Figure 2 .

Figure A5: Effect of learning rate (LR) on models trained from scratch for the [ $\alpha$ /Fe]–[Fe/H] diagram. Conventions follow Figure 3 .

Figure B6: Saliency maps of the output iron abundance with respect to the input spectra in the rest frame. The upper panel shows the median and 1 $\sigma$ dispersion of the normalized spectra for the testing metal-poor subset ([Fe/H]). The middle panel presents the saliency maps (absolute gradients; median and 1 $\sigma$ ) for the residual-head fine-tuning, indicating the sensitivity of the output iron abundance to different spectral regions. The lower panel highlights selected spectral lines with prominent sensitivity, comparing all fine-tuning strategies and the zero-shot results. Lines (from 27 ) corresponding to major spikes are shown in grey (iron) and red (non-iron) in the upper and middle panels, and selectively plotted in the lower panel.

Sample	Selection Criteria	Size	Notes
DESI DR1
Initial cross-match	APOGEE: { STARFLAG , FE_H_FLAG , MG_FE_FLAG } =0 ; SNR $>40$
	DESI: RVS_WARN=0 ; when available, PRIMARY=True	8,104
Training sample	$T_{\rm eff}>4500$ K, 469 metal-poor + 1,600 metal-rich	2,069	For fine-tuning or scratch training
Testing sample	Remaining stars with $T_{\rm eff}>4000$ K	5,539	For testing only
$[\alpha/\text{Fe}]$ –[Fe/H] reference	$T_{\rm eff}>4000$ K after initial cross-match	7,608	For APOGEE [ $\alpha$ /Fe]–[Fe/H] reference

Appendix

Table C3: Summary of DESI DR1 and DESI EDR samples adopted in the main analysis

Figure D7: Comparison of [Fe/H] and [ $\alpha$ /Fe] estimates between the clean, calibrated DESI SP subset and MLP-based models. Plotting conventions follow Figure 2 .

Figure D8: [ $\alpha$ /Fe]–[Fe/H] diagrams for the DESI EDR and the APOGEE reference sample. A clean, calibrated DESI SP subset is used, with common sources shown in the second, third, and fourth columns. Plotting conventions follow Figure 3 .

Cross-Dataset Generalization

Cross-dataset generalization studies whether AI models retain accuracy, calibration, and usefulness on data from sources beyond their training set.

See all 142 papers

Multi-Variable Stellar Parameter Estimation Using Residual Multitask Neural Networks
Jun 11, 2026Bruno Santos Meneses Barreto, Marcio EisencraftParameter Estimation Multi-Task Learning
Escola Polit´ecnica, Universidade de S˜ao Paulo, SP
Spectra as Language: Large Language Models for Scalable Stellar Parameter and Abundance Inference
May 21, 2026Hai-Ling Lu, Yu-Yang Li, Yin-Bi Li +4AI-Assisted Scientific Research
National Astronomical Observatories, Chinese Academy of Sciences, Beijing 100101, China · University of Chinese Academy of Sciences, Beijing 100049, China · School of Astronomy and Space Science, University of Chinese Academy of Sciences, Beijing 100049, China +1
SM-Net: Learning a Continuous Spectral Manifold from Multiple Stellar Libraries
Mar 25, 2026Omar Anwar, Aaron S. G. Robotham, Luca Cortese +1Astrophysics Neural Surrogate Modeling
International Centre for Radio Astronomy Research (ICRAR) The University of Western Australia (UWA), 35 Stirling Hwy, Crawley WA 6009, Australia

Jun 11, 2026J/K move · Enter open · S save

Multi-Variable Stellar Parameter Estimation Using Residual Multitask Neural Networks

Bruno Santos Meneses Barreto, Marcio Eisencraft

Escola Polit´ecnica, Universidade de S˜ao Paulo, SP

Parameter Estimation Multi-Task Learning

Open paper AskPDF arXiv

We present an end-to-end pipeline for estimating stellar parameters from Sloan Digital Sky Survey Data Release 12 spectra using a fully connected multitask neural network with residual blocks, whose hyperparameters are tuned via Bayesian optimization. The preprocessing pipeline includes per-spectrum standardization, RobustScaler normalization of the target variables -- effective temperature

T_{\mathrm{eff}}

, metallicity

[\mathrm{Fe/H}]

, and surface gravity

\log g

-- and data augmentation via Gaussian noise injection. On a held-out test set, the model achieved Mean Absolute Errors (MAE) of

59.76~\mathrm{K}

for

T_{\mathrm{eff}}

0.103~\mathrm{dex}

for

[\mathrm{Fe/H}]

, and

0.130~\mathrm{dex}

for

\log g

. Normalized against the full-scale range of each parameter, these results represent range-normalized errors between

1\%

and

3\%

, achieved with a highly efficient model complexity of approximately 540,000 trainable parameters. These results demonstrate that a compact residual multitask architecture, combined with principled signal preprocessing, provides a parameter-efficient solution for nonlinear parameter estimation in large-scale spectral datasets. In particular, the proposed model achieves competitive performance with substantially lower complexity than deeper neural network baselines.

Abstract

Cross-Dataset Generalization

Explore similar work

Multi-Variable Stellar Parameter Estimation Using Residual Multitask Neural Networks

Spectra as Language: Large Language Models for Scalable Stellar Parameter and Abundance Inference

SM-Net: Learning a Continuous Spectral Manifold from Multiple Stellar Libraries