Generalization from Low- to Moderate-Resolution Spectra with Neural Networks for Stellar Parameter Estimation: A Case Study with DESI
Image generated by Gemini AI
A study addresses the challenge of cross-survey generalization in stellar spectral analysis by using pre-trained multilayer perceptrons (MLPs) to transfer data from LAMOST low-resolution spectra (LRS) to DESI medium-resolution spectra (MRS). The research shows that MLPs pre-trained on LRS perform well even without fine-tuning and improve further with DESI data. While transformer-based embeddings excel for metal-rich stars, they fall short for metal-poor stars compared to MLP-trained models. The choice of fine-tuning strategy varies based on the stellar parameter being analyzed, suggesting that while simple MLPs can effectively generalize across surveys, the potential of spectral foundation models requires additional study.
Neural Networks Enhance Stellar Parameter Estimation Across Surveys
A new study focuses on applying neural networks, specifically multilayer perceptrons (MLPs), to improve the transfer from LAMOST low-resolution spectra (LRS) to DESI medium-resolution spectra (MRS).
The research compares the effectiveness of MLPs trained directly on spectral data against those utilizing embeddings from transformer-based models. Different fine-tuning strategies were also assessed, including residual-head adapters, LoRA, and full fine-tuning.
Key Findings
- MLPs pre-trained on LAMOST LRS demonstrated robust performance even without fine-tuning.
- Modest fine-tuning with DESI spectra improved outcomes across several stellar parameters.
- Transformer-based embeddings outperformed MLPs for iron abundance measurements in metal-rich regimes, while MLPs excelled in metal-poor conditions.
- The choice of fine-tuning strategy significantly influenced results, varying by specific stellar parameters.
Related Topics:
📰 Original Source: https://arxiv.org/abs/2602.15021v1
All rights and credit belong to the original publisher.