Characterization of Gravitational Waves Signals Using Neural Networks

Gravitational wave astronomy has been already a well-established research domain for many years. Moreover, after the detection by LIGO/Virgo collaboration, in 2017, of the first gravitational wave signal emitted during the collision of a binary neutron star system, that was accompanied by the detection of other types of signals coming from the same event, multi-messenger astronomy has claimed its rights more assertively. In this context, it is of great importance in a gravitational wave experiment to have a rapid mechanism of alerting about potential gravitational waves events other observatories capable to detect other types of signals (e.g. in other wavelengths) that are produce by the same event. In this paper, we present the first progress in the development of a neural network algorithm trained to recognize and characterize gravitational wave patterns from signal plus noise data samples. We have implemented two versions of the algorithm, one that classifies the gravitational wave signals into 2 classes, and another one that classifies them into 4 classes, according to the mass ratio of the emitting source. We have obtained promising results, with 100% training and testing accuracy for the 2-class network and approximately 95% for the 4-class network. We conclude that the current version of the neural network algorithm demonstrates the ability of a well-configured and calibrated Bidirectional Long-Short Term Memory software to classify with very high accuracy and in an extremely short time gravitational wave signals, even when they are accompanied by noise. Moreover, the performance obtained with this algorithm qualifies it as a fast method of data analysis and can be used as a low-latency pipeline for gravitational wave observatories like the future LISA Mission.