Learning Representations of Galaxies from Observations and Simulations
speaker: Suchetha Cooray
In recent years, advances in machine learning have allowed astronomers to extract valuable insights from large datasets of observations and simulations. Traditionally, representing information of these datasets has been through summary statistics like the scaling relations, stellar mass functions, and correlation functions. However, we now know that there is plenty of information that may be missed out through the use of summary statistics. Therefore, there is a strong motivation to learn efficient representations made possible through unsupervised machine learning. In this talk, I will present two applications of representation learning on observations and simulations. The first application shows that the emissions of local galaxies can be represented by two parameters that correspond to the galaxy’s evolutionary stage and scale. The second application presents a data-driven model of star formation history (SFH) constructed by finding an efficient representation of simulated SFHs in cosmological simulations. This data-driven SFH model contains the physics included in the simulations and can be used to constrain observed SFHs. I will cover the machine learning techniques used to learn the efficient representations for anyone interested in such methods.