Appendix of the paper entitled "Chasing the formation history of the Galactic metal-poor disk".

This appendix provides details on the estimation of chemical abundances. For the machine learning reference set, we used stars from Li et al. (2022a) and JINAbase Abohalima & Frebel (2018), Fig. A.1 presents the distribution of stellar parameters for the reference set and target stars. Noting systematic biases between the two sources. Fig. A.2 compares atmospheric parameters and chemical abundances for stars common to both sources, and Table A.1 lists parameters for revising abundances from JINAbase Abohalima & Frebel (2018), with details in Section 2.1.1. Table A.2 specifies the wavelength ranges used to estimate Fe, Mg, Ca, and C abundances via template fitting. Fig. A.3 shows machine learning accuracy on training and test sets, while Fig. A.4 illustrates the relationship between errors and prediction accuracy. Fig. A.6, A.7, A.8, and A.9 compare Fe, Mg, Ca, and C abundances measured by machine learning and template fitting, with subplots showing results with and without error filtering. We estimated the atmospheric parameters and chemical abundances of ∼ 100, 000 metal-poor stars using machine learning. For Fe, C, Mg, and Ca, we also conducted measurements through template fitting. To evaluate the accuracy of our results, we compared our values with those from Li et al. (2022b). Fig. A.5 presents the comparison between our machine learning results and those from Li et al. (2022b), while Fig. A.10 shows the comparison of template fitting results with values from Li et al. (2022b).

本附录详述了化学丰度的估算方法。针对机器学习参考集，我们采用了Li等人(2022a)的恒星样本，以及JINAbase数据库中Abohalima与Frebel(2018)的相关恒星数据。图A.1展示了参考集与目标恒星的恒星参数分布，并指出了两个数据源间存在的系统偏差。图A.2对比了两个数据源中共有的恒星的大气参数与化学丰度；表A.1列出了用于修正JINAbase数据库中Abohalima与Frebel(2018)所得丰度值的参数，详细说明参见2.1.1节。表A.2给出了通过模板拟合（template fitting）估算铁(Fe)、镁(Mg)、钙(Ca)与碳(C)丰度所用的波长范围。图A.3展示了机器学习模型在训练集与测试集上的准确率表现，而图A.4则阐明了误差与预测准确率之间的关联。图A.6、A.7、A.8与A.9分别对比了机器学习与模板拟合测得的铁、镁、钙及碳丰度，其子图分别展示了有无误差过滤时的结果。我们通过机器学习方法估算了约10万颗贫金属星的大气参数与化学丰度。针对铁、碳、镁与钙元素，我们同时采用模板拟合方法开展了丰度测量。为评估所得结果的准确性，我们将自身的测量值与Li等人(2022b)的结果进行了对比：图A.5展示了我们的机器学习结果与Li等人(2022b)结果的对比情况，而图A.10则对比了模板拟合结果与Li等人(2022b)的测量值。