Early warning for gravitational wave signals from binary neutron star coalescence using field programmable gate arrays

Abstract

The present thesis is a proof-of-concept study for the possibility for a machine learning-based pipeline to run on an field programmable gate array in order to detect gravitational waves emitted by the early inspiral of binary neutron star mergers. It explores two main neural network models, GregNet and GWaveNet, and their performance at different parts of the inspiral phase of binary neutron star mergers. At a false alarm probability of 1%, they achieve accuracies of 66.81% and 76.22% respectively. It is found that GregNet is 1.5 times faster at inference than GWaveNet on the graphical processing unit and 5.6 times on the central processing unit. Still, the models would be very similar in cost when run as part of a pipeline. The models are successfully adapted to run on the field programmable gate array and quantized. GregNet is compiled to run on the field programmable gate array. The models in general run the fastest in the graphical processing unit. However, the graphical processing unit is also the least energy-efficient and most costly, while the field programmable gate array is the least costly.