Applications of deep learning for gravitational wave physics

Multi-messenger Astrophysics is an emerging multidisciplinary field that demands fast data analysis. On the other hand, deep learning has greatly accelerated signal processing in many fields without sacrificing performance. This work shows that deep learning can be deployed for efficient binary neutron star merger predictions, real-time gravitational wave detections, and gravitational wave denoising. For eccentric compact binary coalescence events, the state-of-the-art early warning system developed in this work is capable of providing up to 2 minutes predictions for the imminent merger events embedded in real advanced LIGO data. A separately developed real-time detection system is able to detect all the gravitational wave events from O2 and O3 data, with a false positive rate of 1 misclassification for every 2.7 days of searched data. As a part of the gravitational wave data analysis process, a denoising neural network is also trained and deployed to remove the noise from the merger events and extract the clean gravitational wave signals. The denoising network is robust against glitch contamination and generalizes well to parameters that are not in the training dataset. The recovered gravitational signals are up to $99\%$ consistent with LIGO results. All the neural networks developed in this work can be deployed to a single graphics processing unit (GPU) without sacrificing performance. The computational cost can be further reduced if the trained neural networks were run on specialized computational hardware that supports quantization. These results show that deep learning provides fast and efficient solutions to big data challenges posed by multi-messenger astrophysics