An Interpretable Hybrid Deep Learning Model for Stellar Parameter Prediction with LAMOST Low-resolution Spectra

作者：Wen-Xuan Zhu, Ming Yang, Sicong Liu, Maosheng Xiang, Ying Liu, Hui-Gen Liu, Linquan Gong, Huan Xie, X Tong · 发表于：The Astrophysical Journal · 年份：2026 · DOI：10.3847/1538-4357/ae7b25 · 研究领域：Stellar, planetary, and galactic studies、Astrophysics and Star Formation Studies、Astronomy and Astrophysical Research

Abstract In recent years, the rapid growth of spectroscopic surveys and advances in deep learning have made stellar parameter estimation increasingly accurate and scalable. However, many existing methods underuse the joint information encoded in multiple spectral lines, particularly weak metal lines, leading to limited analysis of attention shifts and physical interpretability. This study proposes a hybrid prediction framework that integrates a convolutional neural network (CNN), long short-term memory, and a multilayer perceptron (MLP), and validates it with low-resolution LAMOST spectra. The results demonstrate that, compared to CNN, CNN with sequence modeling (CNN-SM) can reduce the mean absolute error (MAE) of effective temperature, surface gravity, and metallicity to 81.190 K, 0.090 dex, and 0.053 dex, respectively. Moreover, the joint CNN-SMRC model, which combines CNN-SM with residual correction, achieves further improvements by leveraging spectral priors to refine the model backbone. The MAEs for the three parameters are further reduced to 56.350 K, 0.087 dex, and 0.049 dex. The MLP feature importance analysis highlights spectral lines and bands with clear physical significance, consistent with astrophysical expectations. This work can support data processing for large-scale sky surveys such as LAMOST and LSST, enhance studies on star formation and evolution, and provide a competitive and interpretable framework for related research.