A New, Data-Driven Era for Precision Cosmology: Measuring the Expansion Rate of the Universe with Machine Learning.

随着新一代的天空观测在未来十年上线，宇宙学即将进入一个大数据的新时代，这势必会给这个领域带来革命性的变化。这些新的实验将提供丰富的数据，以回答诸如暗物质和暗能量的本质等基本问题，但也为解决宇宙学中正在出现的危机打开了一扇新窗口：哈勃常数H0的测量，目前在早期（宇宙微波背景）和晚期（1a型超新星）宇宙探测器之间几乎处于紧张状态。如果得到证实，这种分歧将意味着超越宇宙学标准模型的新物理学，这使得精确确定这个参数成为现代宇宙学实验最重要的目标之一。我的研究项目将利用先进的机器学习方法来释放新一代天文台丰富数据集的力量，利用新一代强透镜类星体的图像来测量H0，从而有可能解决这一危机。这将通过发展复杂和创新的深度学习管道来实现，包括以下三个短期目标：-生产快速，廉价和极其逼真的强透镜类星体观测模拟；-开发一种新的机器学习分析管道，以完全自动化和准确的方式从观察中获得镜头模型和时间延迟；-开发无似然贝叶斯推理框架，以估计预测参数值的校准不确定性。这将以一种自一致的方式整合我们对透镜、类星体物理和宇宙学的最佳理解，这种方式在计算上无法用传统方法进行处理，在即将到来的调查中，可能会以一个数量级增加这门科学可用透镜系统的数量。它将使我们能够充分利用即将到来的新一代天空调查数据的真正潜力。该基金将为机器学习宇宙学这一新兴领域的本科生、硕士和博士生提供培训。在大数据时代，他们的专业知识和工作将由“发现基金”资助，为全球透镜社区提供关键资源和辅助科学产品，开发的工具和产品将开放源代码，供更广泛的社区使用。

With a new generation of sky surveys coming online in the next decade, cosmology is about to enter a new era of big data that is bound to revolutionize this field. These new experiments will provide a wealth of data to answer fundamental questions such as the nature of dark matter and dark energy, but are also opening a new window to solve an emerging crisis in cosmology: the measurement of the Hubble constant, H0, which is now at an almost 4s tension between early (the cosmic microwave background) and late (type 1a supernovae) universe probes. If confirmed, this disagreement would imply new physics beyond the standard model of cosmology, which renders making a precise and accurate determination of this parameter one of the most crucial goals of modern cosmology experiments. My research program will use the power of advanced machine learning methods to unlock the power of the rich datasets from this new generation of observatories to measure H0 using images of a new population of strongly lensed quasars, potentially resolving this crisis. This will be achieved through the development of sophisticated and innovative deep learning pipelines and includes the following three short-term objectives: -The production of fast, inexpensive, and extremely realistic simulations of strongly lensed quasar observations; -The development of a novel machine learning analysis pipeline to obtain lens models and time delays from observations in a completely automated and accurate manner; -The development of a likelihood-free Bayesian inference framework to estimate calibrated uncertainties of predicted parameter values. This will integrate our best understanding of lensing, quasar physics, and cosmology in a self-consistent manner that is not computationally tractable with traditional methods, potentially increasing the number of useable lensing systems for this science in upcoming surveys by an order of magnitude. It will enable us to fully exploit the true potential of the upcoming data from the new generation sky surveys. This grant will provide training for undergraduate, M.Sc., and Ph.D. students in the emerging field of machine learning cosmology. In the era of big data, their expertise and work funded by this Discovery Grant will provide critical resources and ancillary science products for use by the global lensing community, and the tools and products developed will be made open source for the use of the broader community.