Novel Deep Learning Approach to Detecting Binary Black Hole Mergers

Gravitational wave detection has opened up new avenues for exploring and understanding some of the fundamental principles of the universe. The optimal method for detecting modelled gravitational-wave events involves template-based matched filtering and performing a multi-detector coincidence search in the resulting signal-to-noise ratio time series. In recent years, advancements in machine learning and deep learning have led to a flurry of research into using these techniques to replace matched filtering searches and for efficient and robust parameter estimation of the gravitational wave sources. This paper presents a feasibility study for a novel approach to detecting binary black hole gravitational wave signals, which utilizes deep learning techniques on the signal-to-noise ratio time series produced from matched filtering. We show that a deep-learning search can efficiently detect binary black hole gravitational waves from the signal-to-noise ratio time series in simulated Gaussian noise with simulated transient glitches. Furthermore, our search method can outperform a maximum SNR-based matched filtering search on simulated data of the Hanford and Livingston LIGO detectors in the presence of glitches. We further demonstrate that our approach can improve the detection sensitivity for binary black hole mergers at lower masses, relative to a baseline sensitivity of existing search pipelines and deep learning approaches. Lastly, since we are building upon the foundations of a matched filtering search pipeline, we can extract estimates for the signal-to-noise ratio and detector frame chirp mass of a gravitational wave event with similar accuracy as existing pipelines.