BurstDetector: Using Machine Learning to Predict Starbursts in Galaxy Mergers

Abstract

When two galaxies merge together, several phenomena such as starbursts can ensue. Studying such phenomena is crucial to understanding how galaxies and stars interact, with the analysis of star formation rates providing insight into the many mysteries of our universe. Current methods of direct imaging and spectral analysis in analyzing mergers are mostly manual and not automated, oftentimes being prone to error as well. With large datasets emerging online through more and more readily available simulation data, more efficient methods must be developed to study such data. Machine learning techniques can expedite such processes, with this paper aiming to evaluate 3 techniques known for successful image classification in their success with automating the analysis of images from these datasets: Convolutional Neural Networks (CNN), Random Forest Algorithms (RF), and Support Vector Machines (SVM). Trained and cross-validated on image data from the Sloan Digital Sky Survey, our CNN ”BurstDetector” yielded the most success with an accuracy of 92.7% in detecting the occurrence of starbursts, demonstrating that CNNs tend to experience the most success in this image classification task of the 3 models evaluated. BurstDetector can also be run on a multitude of computers regardless of their GPU, making it computationally efficient. A computationally efficient model like BurstDetector is essential to being able to interpret the tremendous amount of data online. The study of the resultant stars forming in merging galaxies through their images can be pivotal to making new discoveries in the field of physics and astronomy, opening the door to revelations in the structure of the universe and even progress with dark matter.