The growth and assembly of galaxies involves many complex processes, which culminate in the diverse collection of galaxies we observe today. One of the best ways of understanding these processes is through “Galactic Paleontology”, which tries to reconstruct the assembly history of nearby galaxies through their surviving “fossils” (which are their present-day surviving stars!). Using this data, we simulate the birth, evolution, and death of many thousands/millions of stars, compare the end result with the stars we observe today, and repeat this process many times for many different evolutionary pathways to see which ones match the observed data better.

For the past few years, astronomers have largely relied on simulation studies and more “ad hoc” approaches to try to compare simulated data with real data, often involving “binning” the data into larger groups. In this project, co-supervised with Prof. Ting Li, we will develop a new, more principled approach based on Inhomogeneous Poisson Point Processes (IPPP) that will allow us to utilize all of the available data. If time/interest permits, we will also try to compare these results with traditional approaches and potentially explore new probabilistic machine learning-driven methods. The main responsibilities of the student will be to review relevant literature, lead coding and data analysis efforts (using simulation studies and/or real data), and meet regularly with me and various collaborators to discuss progress on the project.

Many machine learning (ML) applications rely on having high-quality, labeled training data that is representative of the type of data the ML model will eventually be applied to. However, in many astronomical applications, we have the exact opposite, with observed training data that is substantially biased (often to the brightest, closest, best-measured objects) relative to the underlying populations of interest (the fainter, faraway, noisier objects). To account for these domain mismatch issues, astronomers often resort to various data augmentation strategies that include making the training data “noisier” and supplementing observed data with simulated data from theoretical models. While these broadly address the fundamental problems, they also tend to degrade the performance of the initial ML model. This project will explore new approaches to improve on these data augmentation strategies using state-of-the-art data from the DESI and SDSS-V astronomical surveys, with the goal of having a model that does strictly better on both observed (real) and simulated (theoretical) data under almost all circumstances. The main responsibilities of the student will be to review relevant literature, lead coding and data analysis efforts (using simulation studies and/or real data), and meet regularly with me and various collaborators to discuss progress on the project.