恒星是研究恒星物理和银河系结构等天文核心科学问题的重要探针。大视场光学巡天 计划的兴起及望远镜仪器硬件的飞速发展使得当前恒星光谱和测光数据规模和精度大幅 提高，为恒星参数的高精度估计奠定了坚实的数据基础，同时也对定标方法提出了更高精 度的需求。深度学习方法作为一种数据驱动的算法，能够自动提取数据集特定的深层特 征，为海量光谱、测光等恒星观测量的处理与分析打开了新的思路。利用深度学习方法实 现恒星测光和光谱的高精度定标，以及恒星大气基本参数和元素丰度的高精度估计，是计 算机技术于自然科学研究的重要应用，对人类研究恒星性质和银河系演化意义重大。

针对现有研究中存在的恒星测光和光谱流量定标精度低、恒星参数的可靠测量范 围窄、海量观测数据处理效率低等问题，本文以欧洲的 Gaia 空间测光巡天、西班牙的 Javalambre Photometric Local Universe Survey(J-PLUS)地面测光巡天、中国空间站工程巡 天望远镜(CSST)为例，利用深度学习方法突破了恒星观测量定标和恒星参数估计中的 精度瓶颈。本文主要研究工作和贡献如下:

(1)提出了一种基于多层感知神经网络的测光星等定标方法，能够有效解耦星等间的 相互影响，显著提高星等定标精度，增强模型在物理意义上的可解释性。传统方法通过恒 星基本大气参数限制内禀颜色，但星等之间的耦合作用极大地增加了定标难度。本文在恒 星颜色回归方法的物理规律限制下，首先利用 Landolt 标准星中基于电荷耦合元件(CCD) 的高质量 U、B、V、R、I 测光星等构造测光颜色;其次在多层感知神经网络的隐藏层提 取恒星参数的高维组合特征;然后在训练过程中进行异常观测数据识别和处理以减小模型 带来系统偏差;最后结合物理特征解耦 Gaia 的 3 个波段(G、BP 和 RP)星等。实验结果 表明，相比于之前数据，本文方法实现了定标精度从 1% 到 1 ‰ 的突破。公开发布的星等 修正曲线为恒星参数测量和天文科学问题的研究提供了高质量的测光数据基础。

(2)提出了一种基于代价敏感学习的测光参数测量方法，该方法不再依赖于理论恒星 大气模型，能够精确测量更多精确、可靠的恒星参数。传统的理论恒星大气模型方法计算 量大且结果容易受模型假设影响，随着恒星参数维度的增加和恒星样本在恒星参数空间分 布的不均，恒星参数测量精度的提升变得更加困难。针对以上问题，提出了一种基于代价 敏感学习的多层感知神经网络。该方法在训练得到恒星颜色和恒星参数之间数学关系模型

I

时，根据恒星样本在二维恒星参数空间中分布赋予恒星样本权重，在提升恒星参数测量精 度的同时保证模型结果符合一般物理规律。实验结果表明，该方法不仅能从高质量测光 数据中推断出准确的恒星基本大气参数(有效温度 Teff、表面重力加速度 log g、金属丰度 [Fe/H])，还能预测出多个可靠准确的元素丰度(C、N、Mg、Ca 和 [α/Fe])，公开发布的 J-PLUS DR1 星表为银河系化学动力学分析提供了强有力的数据支撑。

(3)本文提出了一种基于注意力机制和不确定性学习的无缝光谱定标方法，能够融合 多种信噪比(SNR)的观测光谱生成精确、可靠的能谱分布(SED)和相应的预测误差。 在实际观测中，低信噪比的观测光谱占据主导地位，如何增强方法在多种信噪比观测光谱 下的普适性成为一个亟待解决的问题。本文首先利用注意力机制模拟光谱软阈值化降噪过 程，然后分别提取与光谱流量和噪声相关的深层特征，最后在 SED 预测值的不确定性的 引导下实现模型的端到端训练。实验结果表明，该方法对流量标准星的 SED 预测可靠，并 且相应的误差预测值随 Teff、波长和输入光谱的 SNR 呈现符合一般物理规律的趋势变化。 当Teff =6000K时，对于SNR=20、40和80输入光谱，GU通带上SED预测值的典型精 度分别达到 4.2%、2.1% 和 1.5%;当 Teff 升高到 8000 K 时，对应的典型精度分别提高至 1.2%、0.6%和0.5%;对于颜色偏红的GI通带，Teff =6000K时对应的典型精度分别提高 至 0.3%、0.1% 和 0.1%。该方法获取的海量标准星为 CSST 无缝光谱的高精度流量定标奠 定了数据基础。

(4)基于上述深度学习方法，本论文在 Gaia DR3 无缝光谱的恒星参数测量上进行了 应用，不仅能够从多种信噪比的无缝光谱中推断出恒星金属丰度 [Fe/H]，还能够预测出 [Fe/H] 的误差，同时，[Fe/H] 的预测下限进一步延伸至 −4.5。该方法通过注意力机制模块 提取噪声鲁棒的深层特征，并提高对贫金属恒星的关注度，在 [Fe/H] 预测值的不确定性的 引导下推断出目标恒星准确的 [Fe/H] 的值和相应误差。实验结果表明，该方法能够为低分 辨率的无缝光谱确定准确的恒星参数。当 [Fe/H] < −3.0 时，[Fe/H] 的测量精度能够达到 0.27dex。

综上，本文围绕恒星观测量及参数估计问题，针对测光和光谱不同的观测方式和数据 特点，结合深度学习方法的优势与不足，分别提出对应的高精度流量定标和恒星参数测量 方法，为天文观测数据的高效利用和天文科学问题的研究扫清了障碍。本文相关实验结果 较现有方法实现了显著的精度提升，也获得了天文学领域研究人员的认可，较好的实现了 计算机应用技术促进天文学发展的研究目标。

Stars play an important role in the studies of stellar physics and Galactic structure in astron- omy. The rise of large-scale surveys and the rapid development of telescope hardware have greatly improved the number and quality of stellar spectra and photometry. This has laid a solid data foun- dation for high-precision estimation of stellar parameters and has also put forward higher precision requirements for calibration methods. As a data-driven algorithm, deep learning can automatically extract deep features specific to a dataset, opening up new ideas for processing and analyzing mas- sive stellar observable such as spectra, and photometry. Using deep learning methods to achieve high-precision calibration of stellar photometry and spectra, as well as high-precision estimation of basic atmospheric parameters and elemental abundance of stars, is an important application of computer technology in natural science research and has significant implications for human re- search on stars and the evolution of the Milky Way.

In response to the low accuracy of photometric and flux calibration and the narrow range of reliable estimation of stellar parameters in existing research, as well as the low efficiency of processing massive observational data, this paper uses deep learning methods to break through the accuracy bottleneck in stellar observable calibration and stellar parameter estimation. The thesis takes the European Gaia survey, the Spanish Javalambre Photometric Local Universe Survey (J- PLUS) ground-based photometric survey, and the Chinese Space Station Telescope (CSST) as examples. The main research work and contributions of this thesis are as follows:

(1) A photometric magnitude calibration method based on a multilayer perceptron neural net- work is proposed, which can effectively decouple the mutual influence between magnitudes, sig- nificantly improve the accuracy of magnitude calibration, and enhance the interpretability of the model. Traditional methods limit intrinsic color using basic atmospheric parameters of stars, but the coupling effect between magnitudes greatly increases the difficulty of calibration. This pa- per first constructs photometric colors based on charge-coupled device (CCD)-based high-quality UBVRI magnitudes in Landolt standard stars under the physical law constraints of stellar color re-

III

gression methods. Secondly, it extracts high-dimensional combined features of stellar parameters in the hidden layer of multilayer perceptron neural network. Then it identifies and processes abnor- mal observation data during training to decrease systematic bias introduced by the model. Finally, it decouples Gaia’s three bands (G, BP and RP) magnitudes with physical features. Experimental results show that compared with previous data, this method has achieved a breakthrough in cal- ibration accuracy from 1% to 1 ‰. The publicly released magnitude correction curve provides a high-quality photometric data basis for stellar parameter estimation and astronomical science research.

(2) A photometric parameter estimation method based on cost-sensitive learning is proposed. This method no longer relies on theoretical stellar atmospheric models and can accurately estimate more accurate and reliable stellar parameters. Traditional theoretical stellar atmospheric model methods have large computational costs and their results are easily affected by model assumptions. With the increase of stellar parameter dimensions and the uneven distribution of stellar samples in the stellar parameter space, it becomes more difficult to improve the accuracy of stellar parameter estimation. To address these issues, a multilayer perceptron neural network based on cost-sensitive learning is proposed. When the mathematical relationship model between stellar color and stellar parameters is obtained during training, this method assigns weights to stellar samples according to their distribution in the two-dimensional stellar parameter space, thereby improving the accu- racy of stellar parameter measurement while ensuring that the model results comply with general physical laws. Experimental results show that this method can not only estimate accurate ba- sic atmospheric parameters of stars (effective temperature Teff , surface gravity acceleration log g, metallicity [Fe/H]) from high-quality photometric data, but also predict multiple reliable and ac- curate element abundances (C, N, Mg, Ca and [α/Fe]). The publicly released J-PLUS DR1 catalog provides strong data support for galactic chemical dynamics analysis.

(3) A slitless spectra calibration method based on attention mechanism and uncertainty learn- ing is proposed, which can fuse observation spectra with multiple signal-to-noise ratios (SNRs) to generate accurate and reliable spectral energy distributions (SEDs) and corresponding errors. In practical observations, low SNR observation spectra dominate. How to enhance the universality of the method under multiple SNR observation spectra has become an urgent problem to be solved. This paper first uses an attention mechanism to simulate the process of spectral soft threshold denoising, then extracts deep features related to spectral flux and noise respectively, and finally realizes end-to-end training of the model under the guidance of the uncertainty of SED prediction values. Experimental results show that this method can predict reliable SEDs of flux standard

IV

stars, and the corresponding error prediction values show a trend that complies with general phys- ical laws with Teff, wavelength and SNR of input spectra. When Teff =6000 K, for input spectra with SNR=20, 40 and 80 on GU band, the typical accuracy of SED prediction values is 4.2%, 2.1% and 1.5% respectively; when Teff increases to 8000 K, the corresponding typical accuracy is improved to 1.2%, 0.6% and 0.5% respectively; for GI band with reddish color, the corresponding typical accuracy when Teff =6000 K is improved to 0.3%, 0.1% and 0.1%. The massive standard stars obtained by this method lay a data foundation for high-precision flux calibration of slitless spectra in CSST.

(4) Based on the above deep learning methods, this paper applies them to the stellar parameter estimation of slitless spectra in Gaia DR3. It can not only infer the stellar metallicity [Fe/H] from slitless spectra with multiple signal-to-noise ratios (SNRs), but also predict the error of [Fe/H]. At the same time, the predicted [Fe/H] is down to −4.5. This method extracts noise-robust deep features through an attention mechanism module, and increases the focus on low-metallicity stars. Under the guidance of the uncertainty of [Fe/H] prediction values, it infers the accurate value and corresponding error of [Fe/H] for target stars. Experimental results show that this method can determine accurate stellar parameters for low-resolution seamless spectra. When [Fe/H] < −3.0, the precision of [Fe/H] can reach 0.27 dex.

In summary, this paper focuses on the problem of stellar observable calibration and stellar pa- rameter estimation. Based on different data characteristics of photometry and spectroscopy, con- sidering the advantages and disadvantages of deep learning methods, corresponding high-precision flux calibration and stellar parameter estimation methods are proposed respectively, which clears obstacles for the efficient use of astronomical observation data and the study of astronomical scien- tific problems. The relevant experimental results of this paper have achieved significant accuracy improvement compared with existing methods, and have also been recognized by researchers in the field of astronomy. It has achieved the research goal of promoting the development of astronomy through the application of computer technology.