Weighing the stars: Data-driven stellar population modeling for the next-generation sky surveys

由于新一代的天空调查（例如欧几里得，鲁宾/LSST等），预计很快就会有大量的天体物理数据，为回答有关暗物质和星系形成的关键问题提供了前所未有的机会。这些调查将产生超过20万个强引力透镜，比目前的样本大两个数量级。这种透镜是星系中暗物质的理想示踪剂。该项目旨在开发和利用机器学习方法，对透镜星系进行恒星合成建模。恒星种群和强透镜模型的结合将使我们能够分离单个系统中透镜的恒星和暗物质成分，并通过分层推理框架，确定暗物质和星系恒星成分的种群水平属性，为暗物质模型和星系形成场景提供独特的约束。引力透镜追踪物质（暗物质和重子的组合）在透镜星系中的分布，通过它们在遥远源的图像中引起的引力扭曲。分离这些星系的暗物质和恒星成分是检验暗物质模型预测和理解星系形成和演化中的重子物理过程的关键。这是通过恒星综合建模完成的，这是一种将模型恒星光谱组合起来以拟合一组宽带图像的过程。虽然由于各种假设和简并，这些模型仍然存在偏差和不确定性，但最近的研究表明，通过将恒星合成的个体测量与强透镜和弱透镜相结合，分别探测星系的内部和外部，对于大量强透镜系统，有可能在统计上推断出暗物质性质和恒星成分的无偏测量。允许对暗物质和星系形成模型进行可靠的测试。使用传统的最大似然建模方法，对即将进行的调查中的海量数据（甚至是当前可用的数据）执行此练习是难以处理的。然而，最近的进展表明，机器学习可以将镜头建模的过程加快1000万倍以上，从而使计算上棘手的镜头建模问题在几分钟内得到解决。我们的团队目前正在构建方法和管道来实现这一点。我们将通过获得恒星质量和建立一个层次推理框架来扩展这些努力，该框架将星系恒星群和透镜参数的测量与无与伦比的大小样本相结合，以便直接探测重子质量分数的时间演化，内/外星系质量比，以及质量吸积的环境依赖性。在预期的全天调查中，欧几里得和鲁宾天文台应该在2023年和2024年开始运作，这使得这个项目最及时。最近发射的詹姆斯·韦伯太空望远镜也将提供恒星种群随时间演变的独特特征，这是我们项目的必要投入。我们的跨学科团队由强+弱引力透镜分析、恒星群建模和机器学习方面的专家组成，在这个令人兴奋的项目的各个方面形成了独特的专家合作。

Remarkable volumes of astrophysical data are expected soon thanks to a new generation of sky surveys (e.g. Euclid, Rubin/LSST, etc.), providing unprecedented opportunities to answer key questions about dark matter and galaxy formation. These surveys will yield more than 200,000 strong gravitational lenses, two orders of magnitude greater than current samples. Such lenses are ideal tracers of dark matter in galaxies. This project aims to develop and exploit machine learning methods that perform stellar synthesis modeling of lensing galaxies. The combination of stellar population and strong lens modeling will allow us to separate the stellar and dark matter components of lenses in individual systems and, through a hierarchical inference framework, to determine population-level properties of dark matter and stellar components of galaxies, providing unique constraints on dark matter models and galaxy formation scenarios.Gravitational lensing traces the distribution of matter (dark matter and baryons combined) in lensing galaxies through the gravitational distortions they cause in images of distant sources. Separating the dark and stellar components of these galaxies is the key to testing predictions of dark matter models and understanding the baryonic physics processes in galaxy formation and evolution. This is done using stellar synthesis modeling, a procedure by which model stellar spectra are combined to fit a set of broadband images. While these models remain biased and uncertain due to various assumptions and degeneracies, recent studies have demonstrated that by combining the individual measurements of stellar synthesis with strong and weak lensing, which probe the inner and outer parts of a galaxy respectively, for a large population of strong lensing systems, it becomes possible to statistically infer unbiased measurements of dark matter properties and stellar components, allowing robust tests of dark matter and galaxy formation models. Performing this exercise for the monumental volumes of data from upcoming surveys (and even for currently available data) is intractable with traditional maximum-likelihood modeling approaches. However, recent advances have shown that machine learning can accelerate the process of lens modeling by more than 10 million times, allowing the computationally intractable lens modeling problem to be solved in minutes. Our team is currently building methods and pipelines to do just that. We will expand upon these efforts by obtaining stellar masses and building a hierarchical inference framework that combines the measurements of galaxy stellar populations and lensing parameters to samples of unparalleled size in order to directly probe the time evolution of the baryonic mass fraction, the inner/outer galaxy mass ratio, and the environmental dependence of mass accretion. Among the anticipated all-sky surveys, the Euclid and Rubin Observatories should begin operations in 2023 and 2024, making this project most timely. The recently launched James Webb Space Telescope will also provide a unique characterization of the evolution of stellar populations with time, a necessary input for our project. Our interdisciplinary team is composed of experts in strong+weak gravitational lensing analysis, stellar population modeling, and machine learning, forming a unique collaboration of experts for each aspect of this exciting project.