利用深度学习重构本地宇宙结构演化

LCDM cosmological model is supported by a lot of observational evidences and almost becomes the standard cosmological model. However, we are still not very clear on some issues on understanding these observational data, e.g. cosmic variance; bias between observed galaxy distribution and dark matter background; the formation history of observed various structures. So it is very essential to reconstruct the evolution of the observed local universe to advance our understanding on these issues. This proposal will adopt the deep learning method, which is completely different from any traditional methods, to address this problem. With multilayer neural network (MNN), we will find out the mapping between the observed galaxy and total mass distribution, the mapping from the final obvserved mass distribution to their initial condition. With the proper reconstructed initial condition of the observed local universe( e.g. SDSS, 2MASS), we could reconstruct the evolution history of local universe by virtue of high resolution simulations. The MNN method is searching the link between the initial condition and final state directly, so it is much reliable to get the formation history of the structures in the observed universe. Also with more input information for MNN, it is not difficult to make the reconstruction better than any traditional methods at small scales given without overfit problems. The reconstructed mass distribution is very useful on understanding the total mass distribution in the observed universe, and the formation history of any structures inside.

LCDM宇宙学模型获得大量观测支持从而成为标准宇宙学模型。但为了更好的理解这些观测数据，比如大尺度观测上的宇宙方差（cosmic variance），星系分布与背景物质密度偏差（bias）,以及理解观测到的本地宇宙中的各类结构的形成历史，我们都有必要重构观测到的宇宙的演化图景。本申请将采用和传统方法完全不一样的深度学习的方法来完成重构。首先利用多层神经网络的方法，寻找观测的星系分布(比如SDSS，2MASS)和背景总体物质（暗物质和重子物质）的联系；再通过深度学习找到从当前物质分布到原初物质分布的映射关系从而得到观测宇宙的原初条件；继而对这一原初条件进行高精度数值模拟重构本地宇宙的演化历史。深度学习是直接获得演化开始和结果的映射，故而可以更可靠的获得结构形成历史。同时可以通过增多对深度学习的信息输入，我们也可能可以获得比以前传统方法在更小尺度上的重构。重构出来的本地宇宙演化图景将有诸多应用

国内外新一代星系巡天都正在逐步开展，如果从星系巡天获得更多有用信息成为当前整个领域的重要挑战之一，本项目另辟蹊径，尝试利用新工具---人工智能技术从全新的途径完成该项目。本项目完成了所有项目规划，基本实现了科学目标：完成了大规模数值模拟Indra.全新有自主知识产权的星系形成与演化模型GABE（程序约2.5万行），从星系巡天到宇宙密度场重构的三个最重要关键模块的建模：非线性演化，红移畸变，偏袒因子。其中对宇宙密度场的重构的精度均与传统方法相当，或者更优。重要成果发表相关论文10余篇。基于目前的成果，我们将能利用已有的星系巡天或者新一代巡天的观测数据完成对宇宙物质密度场的重构，从而对宇宙学模型给出更精确的限制，并理解观测到的天体或者结构的形成演化历史。