Deep unsupervised machine learning approaches for galaxy evolution studies

点击翻译按钮获取中文摘要

Astronomical observations present essential information for the study of galaxy formation and evolution. This information is usually delivered in spectroscopic or photometric form, each providing valuable insights. Examining a galaxy's spectra reveals crucial information about its properties and internal processes. In other words, a considerable volume of information about galaxies is encoded in their spectra. Various methods can be used to infer different physical parameters, such as the metallicity of galaxies, their stellar masses, and star formation rates. While spectroscopic data provide us with invaluable and detailed information about galaxies, photometric records (in different filters) also present valuable knowledge about the morphological characteristics of galaxies. Recent technological advances in Integral Field Units (IFUs) are ideal for exploring spatially extended sources. These surveys generally provide two-dimensional maps in many different physical parameters, informing us of galactic history. An example of such a survey, and the data to be used in this proposal, is Mapping Nearby Galaxies at the Apache Point Observatory survey (MaNGA). Since these spectroscopic and photometric data are so different, there isn't a classical high-throughput way of combining them. To do this, and take the next step in our understanding of galaxy evolution, we need to use Machine Learning (ML) techniques. Unsupervised methods in ML aim to infer structures from the input data with minimal assumptions. These methods can discover hidden relationships and patterns within the data by abstracting the data to a compressed informative representation called latent data. In addition, the possibility to combine multiple deep unsupervised methods further enhances the scalability and overall performance. My research proposal focuses on utilizing latent data from available photometric and spectroscopic data from the MaNGA survey. MaNGA gives us very informative data, but the trade-off is that it is very complex. For example, it is challenging to compare galaxies using spectra alone when they are sampled in a complicated way. The number of samples from two galaxies can differ by an order of magnitude. Spectroscopic and photometric data for the same galaxies are available. By stitching together these latent representations of two data sets quantifying different characteristics, we will generate more robust summaries of MaNGA galaxies. In this proposal, we're continuing to develop unsupervised ML techniques applied to astronomical data. While building expertise with these methods in new researchers, we will also be contributing to the astronomy and astrophysics research communities by unbiased clustering of galaxies based on internal processes. Summarizing galaxies using these data-driven methods will provide a foundation for comparing different data sets, including simulations, to our observed universe.