Advancing Gravitational Wave Science with the Help of Machine Learning

Last year, I had the opportunity to work with gravitational wave researchers on a project that involved using machine learning techniques for gravitational wave science. I will be discussing the background of the project, what we initially set out to do, and the glitch classification model that we created. Lastly, I’ll talk about the generation of synthetic noise or glitch images using generative adversarial networks.

Gravitational waves were first predicted by Albert Einstein in 1916 on the basis of general relativity theory and were confirmed by observations in September of 2015. The observations were made by the LIGO observatories and matched the signal that would be emitted by the merger of two black holes. Gravitational waves are basically distortions in space-time caused by the inward spiralling motion and merger of two black holes.

OzGrav is the ARC Centre of Excellence for Gravitational Wave Discovery in Australia. Towards the end of 2021, Eliiza formed a partnership with OzGrav with the goal of enhancing machine learning and the application of artificial intelligence in gravitational wave astronomy. Our project focused on carrying out glitch classification using gravitational wave data, and synthetic image creation using generative adversarial networks (GANs).

Glitches are noise transients in the data from the LIGO observatory that are caused by anything from human activity to weather effects. Glitches come at a significant rate and can mimic gravitational wave signals. To address this issue, the Gravity Spy project was started as an effort to classify the glitches into morphological families using machine learning algorithms and the work of citizen scientists. Once we are able to classify these glitches, we might be able to identify the root cause of glitches and help with mitigation efforts.

In this Brown Bag I summarise the work we did with OzGrav, how we used the Gravity Spy dataset and built a convolutional neural network to morphologically classify these glitches, how we addressed the class imbalance issue in the data, and I discuss the results of synthetic glitch image creation with GANs.