Exploring the Milky Way with Deep Learning

Abstract

Understanding the formation and evolution of the Milky Way Galaxy demands multi-dimensional measurements, including kinematics, chemical abundances, and ages, from a vast collection of stars distributed across different parts of the Galaxy. Stars act as fossil records in Galactic archaeology, where their current properties provide clues about past events such as mergers and chemical enrichment. Thus, inferring stellar properties from observations offer insights into major events that occurred during the early evolution of the Galaxy. We are currently in a golden era of big data and Galactic surveys, with projects like SDSS-V's Milky Way Mapper, Gaia, TESS, and other large-scale surveys providing diverse information about stars across a significant volume of the Galaxy. Given the vastness of these datasets, efficient and flexible analysis methods are essential to address various scientific questions. Training data-driven models using machine learning on these datasets, is becoming increasingly popular in data-driven astronomy. Deep learning, a subset of machine learning, is particularly flexible and versatile, capable of adapting to the needs of astronomical data and running efficiently on graphics processing units with large datasets. In this thesis, we apply a variety of deep learning methods to datasets from Galactic surveys, including APOGEE and Gaia. We measure basic parameters of the Galaxy such as the distance to the Galactic center using inferred spectro-photometric distances from trained models. Subsequently, we determine stellar spectroscopic ages with minimal chemical contamination trained with spectra and light-curves. We also leverage recent advancements in large-scale foundation models for natural language processing to demonstrate that with similar methodologies, using Transformers and diffusion models, foundation models can be built for floating-point astronomical data. Our ultimate goal is to develop large-scale foundation models that can be applied to various types of astronomical data in a general manner to perform a wide range of tasks, thus advancing our understanding of the Milky Way Galaxy.