Exploring Gravitational Waves Recordings with Machine Learning Techniques

Abstract

The study of Gravitational Waves (GWs) opened a new window of possibilities to improve our understanding of the Universe. GWs provide suitable astronomical messengers for studying events that were not possible before through electromagnetic radiation, or in other cases complementing their observations. Ground-based interferometers like LIGO have been recording multiple GW events since the first detections in 2015. Despite the success of Earth-based observatories, the space limitations and noise sources on Earth point toward the need of building a spaceborne interferometer. The Laser Interferometer Space Antenna (LISA) is a planned project that will provide us with such a detector and will allow gaining access to lower frequency bands and more types of GW sources. To make the most out of LISA’s strengths, it is important to identify and develop alternative data analysis tools which are more appropriate for low latency searches of GWs than the current ones in use. Machine Learning techniques are a promising candidate since they can provide high accuracies, higher speeds, and a lower computational cost. Therefore, they can be used for the development of Low Latency Detectors (LLD) of GWs, which will be used to analyze the LISA recordings. I propose to build a prototype LLD by using a Sliding Window Algorithm, which makes use of Convolutional Neural Networks (CNNs) as its classification mechanism. To implement the LLD, I first create datasets composed of synthetic GW recordings of two different GW source types: Galactic Binaries (GBs) and Merging Blackhole Binaries (MBHBs). Then, I transform these recordings originally represented only in the time domain, into the frequency domain, and the time-frequency domain and train two different ML architectures (CNNs and Fully-Connected Neural Networks) using both the original and the transformed data. A performance evaluation is done to select the best combination of ML architecture and domain representation for solving the detection task. The chosen combination is then used as the classifier mechanism of the LLD acting in windows of five days duration. The LLD is tested on one-year-long recordings with different levels of noise. The analysis suggests that the time-frequency domain representations offer the most promising results for detecting both types of sources (GBs and MBHBs) reaching high accuracies in recordings with low to moderate signal-to-noise ratio (SNR).