| 其他摘要 | The stellar spectrum can reflect the star's surface temperature, surface gravity, mass, radius, luminosity, chemical composition, and so on. The profile of the star spectrum mainly depends on the physical properties and chemical composition of the star. Therefore, the difference in star spectrum types also reflects the properties of the star. Until now, differences and research on the nature of stars have been obtained from spectra. Measurement of stellar atmosphere parameters is the basis for scientific research using stellar spectra. The China Space Station Telescope (CSST), scheduled to be launched in 2024, will provide researchers with slitless spectra of hundreds of millions of stars within a 10-year survey. Compared with manual processing, machine learning has higher efficiency and stability in processing massive data. With the progress of the CSST large-scale sky survey project, hundreds of millions of slitless spectral data will be produced, providing a basis for us to measure stellar atmosphere parameters and quickly classify observation targets. Therefore, my dissertation revolves around the measurement of atmospheric parameters of early-type stars based on CSST slitless spectra. Based on the design indicators of the CSST slitless spectrum and the machine learning algorithm Stellar LAbel Machine (SLAM), we studied the stellar parameters of early-type stars (effective temperature Teff >15000 K). We used the Potsdam-Wolf-Rayet (POWR) synthetic spectral library as training samples. We used both the POWR simulated spectral library and the Next Generation Spectrum Library (NGSL) observed by the Hubble Space Telescope (HST) to test the reliability of the machine learning results. In the ideal case of a full-wavelength spectrum without any noise, the average relative deviation (after taking the absolute value) of the effective temperature Teff is 800 K (2.7%), and the deviation of surface gravity log g is 0.11 c.g.s (3.5%). Using spectra from the POWR library as samples, we divided the samples into three parts with different resolutions (2550 A to 4050 A (resolution R = 287); 4050 A to 6300 A (R = 232); 6300 A to 10000 A (R = 207)) according to different bands of CSST's slitless spectrum and simulated the actual observation situation by adding the effects of radial velocity RV and interstellar extinction AV to the testing sample for 4-fold cross-validation. We obtained the test accuracy of the model for effective temperature Teff and surface gravity log g under different conditions, and gave the measurement error. We also considered possible pixel shifts in the actual observation equipment and studied the impact on the accuracy of the test results when slight shifts in the spectrum occurred in the testing sample. We find that at the shift of 5 A, the average relative deviation of Teff is 1100 K (3.6%), and the deviation of log g is 0.13 c.g.s (4.2%); in the case of shifting 10 A, the average relative deviation of Teff is 1300 K (4.3%), the deviation of log g is 0.15 c.g.s (5.1%). In addition, we also used the observed spectra data of NGSL as the testing sample to obtain the measurement accuracy of Teff and log g, and compared them with the parameters, which previous researchers measured by the classical method. The deviation of stellar parameters was less than 4000 K (15%). These results show that using a machine learning model based on the SLAM method, we can obtain relatively accurate atmospheric parameters of early-type stars from the massive data of the CSST Slitless Spectroscopic Survey. |
修改评论