Se-ResNet+SVM Model: An Effective Method of Searching for Hot Subdwarfs from LAMOST

Abstract

This paper presents a robust neural network approach for identifying hot subdwarfs. Our method leveraged the Squeeze-and-Excitation Residual Network to extract abstract features, which were combined with experience features to create hybrid features. These hybrid features were then classified using a support vector machine. To enhance accuracy, we employed a two-stage procedure. In the first stage, a binary classification model was constructed to distinguish hot subdwarfs, achieving a precision of 98.55% on the test set. In the second stage, a four-class classification model was employed to further refine the candidates, achieving a precision of 91.75% on the test set. Using the binary classification model, we classified 333,534 spectra from LAMOST DR8, resulting in a catalog of 3086 hot subdwarf candidates. Subsequently, the four-class classification model was applied to filter these candidates further. When applying thresholds of 0.5 and 0.9, we identified 2132 and 1247 candidates, respectively. Among these candidates, we visually inspected their spectra and identified 58 and 30 new hot subdwarfs, respectively, resulting in a precision of 82.04% and 88.21% for these discoveries. Furthermore, we evaluated the 3086 candidates obtained in the first stage and identified 168 new hot subdwarfs, achieving an overall precision of 62.54%. Lastly, we trained a Squeeze-and-Excitation regression model with mean absolute error values of 3009 K for T _eff , 0.20 dex for log g , and 0.42 dex for log( n He/ n H). Using this model, we predicted the atmospheric parameters of these 168 newly discovered hot subdwarfs.

Keywords

• Stellar classification
• A subdwarf stars
• Convolutional neural networks
• Stellar types