基于深度学习技术的恒星大气物理参数自动估计

Stellar Atmospheric Parameterization Based on Deep Learning

深度学习是当前机器学习、模式识别和人工智能领域中的一项热点研究技术,非常适用于处理复杂的大规模数据.基于深度学习理论构建了一个5层的栈式自编码深度神经网络,对恒星大气物理参数进行自动估计,网络各层的节点数分别为3821-500-100-50-1.使用美国大型巡天项目Sloan发布的Sloan Digital Sky Survey(SDSS)实测光谱以及由Kurucz的New Opacity Distribution Function(NEWODF)模型得到的理论光谱进行了实验验证,对有效温度(Teff)、表面重力加速度(lg g)和金属丰度([Fe/H])3个物理参数进行了自动估计.结果表明,栈式自编码深度神经网络的估计精度较好,其中在SDSS数据上的平均绝对误差分别为:79.95(Teff/K),0.0058(lg(Teff/K)),0.1706(lg(g/(cm·s^(-2)))),0.1294 dex([Fe/H]);在理论数据上的平均绝对误差分别是:15.34(Teff/K),0.0011(lg(Teff/K)),0.0214(lg(g/(cm·s^(-2)))),0.0121 dex([Fe/H]).

Deep learning is learning, pattern recognition, a typical learning method and artificial intelligence. widely studied in machine This work investigates the stellar atmospheric parameterization problem by constructing a deep neural network with five layers. The proposed scheme is evaluated on both real spectra from Sloan Digital Sky Survey （SDSS） and the theoretic spectra computed with Kurucz＇s New Opacity Distribution Function （NEWODF） model. On the SDSS spectra, the mean absolute errors （MAEs） are 79.95 for the effective temperature （Teff/K）, 0.0058 for lg （Teff/K）, 0.1706 for surface gravity （lg （g/（cm· s^-2）））, and 0.1294 dex for metallicity （[Fe/H]）, respectively; On the theoretic spectra, the MAEs are 15.34 for Teff/K, 0.0011 for lg （Teff/K）, 0.0214 for lg （g/（cm· s^-2））, and 0.0121 dex for [Fe/H], respectively.