发表机构
University of Toronto; Dunlap Institute for Astronomy & Astrophysics; David A. Dunlap Department of Astronomy & Astrophysics; Department of Statistical Sciences; Data Sciences Institute（多伦多大学; 邓普天文学与天体物理研究所; 大卫·A·邓普普天文学与天体物理系; 统计科学系; 数据科学研究所）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
本文提出flexAE框架，结合自动编码器和分类器，基于宽带测光识别双星，应用于杜鹃座47，分类准确率0.85，估计双星比例为2.2%至7.5%，适用于大规模巡天。
AI 中文摘要
双星系统极为普遍，并影响天体物理学的许多领域，如恒星演化和星团动力学。在这项工作中，我们引入了flexAE，一个将经典自动编码器架构与专用分类器组件相结合的框架。它旨在基于宽带光谱能量分布区分单星和双星系统，同时利用所有可用的测光数据并考虑观测不确定性。我们通过使用来自维拉·C·鲁宾天文台的测光数据，识别杜鹃座47球状星团（NGC 104）外围的未解析主序双星，展示了flexAE的能力。我们在由恒星大气模型生成的单星和双星系统模拟样本上训练模型。该模型准确重建了测光输入特征，并在模拟测试集上达到了0.85的分类准确率。然后，我们将训练好的模型应用于鲁宾DP1中发现的1,424个具有可靠$gri$测光的星团成员样本，其中32个被模型识别为未解析主序双星。这对应于观测到的双星比例为$2.2^{+0.5}_{-0.3}$%。我们估计内在主序双星比例介于$3.7^{+0.8}_{-0.5}$%和$7.5^{+1.5}_{-1.1}$%之间。双星比例在距星团中心18-36角分（5.7-11.4半光半径）范围内保持恒定。我们强调flexAE可以轻松适应其他用例。这使得flexAE非常适合即将到来的宽视场巡天（如鲁宾LSST）中的大规模双星分类。代码可在以下https URL公开获取。
英文摘要
Binary systems are extremely common and influence many areas of astrophysics, such as stellar evolution and cluster dynamics. In this work, we introduce flexAE, a framework that combines a classical autoencoder architecture with a dedicated classifier component. It is designed to distinguish between single stars and binary systems based on their broadband spectral energy distributions, using all available photometry simultaneously while accounting for observational uncertainties. We demonstrate the capability of flexAE by using it to identify unresolved main sequence binaries in the outskirts of the 47 Tucanae globular cluster (NGC 104) with photometric data from the Vera C. Rubin Observatory. We train the model on a simulated sample of single star and binary systems created with stellar atmosphere models. The model accurately reconstructs the photometric input features and achieves a classification accuracy of 0.85 on the simulated test set. We then apply the trained model to a sample of 1,424 cluster members found in Rubin DP1 with reliable $gri$ photometry, of which 32 are identified by the model as unresolved main sequence binaries. This corresponds to an observed binary fraction of $2.2^{+0.5}_{-0.3}$%. We estimate that the intrinsic main sequence binary fraction lies between $3.7^{+0.8}_{-0.5}$% and $7.5^{+1.5}_{-1.1}$%. The binary fraction stays constant between 18-36 arcmin (5.7-11.4 half-light radii) from the cluster center. We highlight that flexAE can be easily adapted for other use cases. This makes flexAE well-suited for large-scale binary classification in upcoming wide-field surveys, such as Rubin LSST. The code is publicly available at https://github.com/tobiasgeron/flexAE.
Comments28 pages, 28 figures, 2 tables
