Optimizing Gravitational Wave Detection Using Topological Data Analysis

Ms. Jillian Cervantes
Undergraduate
University of Wisconsin – Milwaukee

The interactions between black holes create decaying periodic “chirp” signals known as gravitational waves (GW). The detection of GW is an important problem in physics, but proves difficult as the signals are embedded in background noise. In 2019, Bresten and Jung determined that using persistence vectors as topological features for input to a convolutional neural network (CNN) results in a better classifier than the traditional CNN method. Our aim is to determine whether a CNN is necessary, or if better topological features with a simpler classifier can achieve similar results. We evaluate persistence vectors in comparison with other topological features, including persistence landscapes, persistence images, and persistence entropy. We determine that the persistence vectors feature extraction method with a support vector classifier achieves the best results without using a CNN, at 0.943 AUC. This classifier is much easier to train and interpret than a CNN, and is competitive even with a significantly smaller sample size.