Automating Galaxy Classification with Unsupervised Machine Learning

Project Description:

Upcoming telescopes such as the Square Kilometre Array (SKA) and the Vera C. Rubin Observatory, along with precursor instruments like MeerKAT and DECam, are revolutionising astronomical analysis and discovery. As data volumes surge, efficient methods for galaxy classification and rapid identification of rare sources are increasingly critical. Traditional approaches often rely on time-consuming human labelling, underscoring the need for automated solutions. This project builds on Mohale & Lochner (2024) to develop and optimise the clustering algorithm Bayesian Gaussian Mixture Models (BGMM) for both optical and radio data. BGMMs can automatically group similar sources, accelerating training set creation, identifying morphologically complex groups, and revealing new patterns in large datasets. The student will create metrics to evaluate cluster stability, assess variability, and determine the optimal number of clusters. A key innovation is preparing static clustering methods for integration into a cyber-machine interface, allowing humans (including citizen scientists) to focus on labelling ambiguous sources. The project will assess the reliability of BGMM probability estimates against known classes, paving the way for novel unsupervised active learning with next-generation telescopes.

Research Area:

Astronomy

Project Level:

Masters

This Project Is Offered At The Following Node(s):

(UCT)

Special Requirements:

Good programming skills are critical for this project, primarily in python. Experience with machine learning is advantageous, but not essential.