Galaxy Evolution - a multicomponent machine-learning model

科学目标是通过应用为机器学习和人工智能开发的新颖统计技术来了解星系的形成和演化，以建立新的多波长星系数据库。星系外星系群的统计特征对于解决星系和AGN的演化至关重要，无论是对不同时期星系群的经验比较还是对理论模型的对抗。目前从深度多波长调查中获得的大量数据、提出的问题以及解决这些问题的理论模型都是丰富而复杂的。然而，自20世纪70年代以来，所使用的统计度量（光度函数、模板光谱能量分布、2点相关函数和特设缩放关系）几乎没有变化。这些过时的衡量标准加深了我们的偏见，限制了我们的理解。在这个项目中，你将采用一种激进的新方法。利用机器学习技术，我们将在赫歇尔河外遗产项目（HELP）中建立一个巨大的多波长目录和地图数据的概率生成模型。该模型将在有限且定义良好的先验假设下，提供可观测值的稳健概率描述。你将用它来描述星系群的关键发射成分同时在所有波长和它们之间的概率关系。你将专注于恒星形成和AGN组件的理解已经特别有限的特设隔离和分类。完整的后验概率将被充分表征，为我们提供数据的紧凑“描述”。这允许你开发一个工具，可以用来生成合成数据集，与约束数据一致，然后可以用来测试任何物理驱动的星系演化模型。赫歇尔河外遗产项目（HELP https://herschel.sussex.ac.uk/）是一个欧洲资助的项目，将于2018年6月完成。HELP结合了所有世界领先的地面和空间天文台的数据。这些数据是通过雄心勃勃的调查，在不同波长下绘制了超过1000平方度的天空。HELP通过链接和均质化数据，提供数据选择函数的高级描述，获得新的物理量，如红移，恒星质量和恒星形成率，以及访问这些数据的新工具，增加了价值。HELP的一个关键发展是XID+工具，它使用贝叶斯推理对来自Herschel的低分辨率数据进行分层概率模型，使用来自其他设施的高分辨率数据。这个项目将利用来自HELP的数据，特别是利用和扩展XID+框架。

The scientific objective is to understand the formation and evolution of galaxies by applying novel statistical techniques, developed for machine-learning and artificial intelligence, to a new multi-wavelength data-base of galaxies. The statistical characterisation of the extra-galactic populations is fundamental to address the evolution of galaxies and AGN, both for empirical comparisons of populations at different epochs and for confronting theoretical models. The extraordinary wealth of data now available from deep multi-wavelength surveys, the questions being posed, and the theoretical models that address them, are rich and complex. However, the statistical measures being used (luminosity functions, template spectral energy distributions, 2-point correlation functions and ad-hoc scaling relations) have changed little since the 1970s. These outdated measures entrench our prejudices and limit our understanding. In this project you will adopt a radical, new approach. Using techniques from machine learning we will build a probabilistic generative model of the vast multi-wavelength catalogue and map data within the Herschel Extragalactic Legacy Project (HELP). This model will provide a robust probabilistic description of the observables, with limited and well defined prior assumptions. You will use this to characterise the key emission components of galaxy populations simultaneously at all wavelengths and the probabilistic relations between them. You will focus on the star forming and AGN components where understanding has been particularly limited by ad-hoc segregation and classification. The full posterior probabilities will be fully characterised, providing us with a compact "description" of the data. This then allows you to develop a tool that can be used to generate synthetic data sets, consistent with the constraining data, that can then be used to test any physically motivated model of galaxy evolution. The Herschel Extragalactic Legacy Project (HELP https://herschel.sussex.ac.uk/) is a European funded project due to complete in June 2018. HELP combines data from all the world-leading terrestrial and space observatories. These data were taken by ambitious surveys charting over 1000 square degrees of the sky at different wavelengths. HELP has added value through e.g. linking and homogenising the data, by providing high-level descriptions of the selection functions of the data, deriving new physical quantities like redshifts, stellar masses and star formation rates and new tools to access these. A key development in HELP has been the XID+ tool which uses Bayesian Inference for a hierarchical probabilistic model of the low-resolution data from e.g. Herschel using the higher resolution data from other facilities. This project will exploit the data from HELP and in particular exploit, and extend, the XID+ framework.