Deep Learning Techniques to make Gravitational Wave Detections from Weak Time-Series Data

Abstract

Gravitational waves are ripples in the space time fabric when high energy events such as black hole mergers or neutron star collisions take place. The first Gravitational Wave (GW) detection (GW150914) was made by the Laser Interferometer Gravitational-wave Observatory (LIGO) and Virgo Collaboration on September 14, 2015. Furthermore, the proof of the existence of GWs had countless implications from Stellar Evolution to General Relativity. Gravitational waves detection requires multiple filters and the filtered data has to be studied intensively to come to conclusions on whether the data is a just a glitch or an actual gravitational wave detection. However, with the use of Deep Learning the process is simplified heavily, as it reduces the level of filtering greatly, and the output is more definitive, even though the model produces a probabilistic result. Our technique, Deep Learning, utilizes a different implementation of a one-dimensional convolutional neural network (CNN). The model is trained by a composite of real LIGO noise, and injections of GW waveform templates. The CNN effectively uses classification to differentiate weak GW time series from non-gaussian noise from glitches in the LIGO data stream. In addition, we are the first study to utilize fine-tuning as a means to train the model with a second pass of data, while maintaining all the learned features from the initial training iteration. This enables our model to have a sensitivity of 100%, higher than all prior studies in this field, when making real-time detections of GWs at an extremely low Signal-to-noise ratios (SNR), while still being less computationally expensive. This sensitivity, in part, is also achieved through the use of deep signal manifolds from both the Hanford and Livingston detectors, which enable the neural network to be responsive to false positives.