Rapid calculation of the signal-to-noise ratio of gravitational-wave sources using artificial neural networks
Γρήγορος υπολογισμός του signal-to-noise ratio των πηγών βαρυτικών κυμάτων χρησιμοποιώντας τεχνητά νευρωνικά δίκτυα
![Thumbnail](/xmlui/bitstream/handle/unipi/15239/Tsioulis_21054.pdf.jpg?sequence=9&isAllowed=y)
Master Thesis
Author
Τσιούλης, Ιωάννης
Tsioulis, Ioannis
Date
2023-02View/ Open
Keywords
Deep learning ; Machine learning ; Astronomy ; Gravitational-waves ; Artificial neural networksAbstract
Using interferometric gravitational wave detectors, we can now observe the dark side of the
universe. Already, nearly 90 binary black hole systems have been detected. To measure the
source parameters of a detected system, one needs to perform matched filtering, using a large
number of templates (of order hundreds of thousand). In such parameter estimation
calculations it is also useful to know the optimal signal-to-noise ratio of an individual model
waveform. In this thesis, we train an artificial neural network on a random sample of one million
theoretical waveforms of binary black hole systems with random spins and achieve an accuracy
of 97% in predicting the signal-to-noise ratio. The neural network evaluates the results orders
of magnitude faster than the original calculation. We show the results of the optimization of
different hyperparameters with a grid search and with selective searches. Finally, we show that
the logarithm of the accuracy is linearly related to the logarithm of the number of points in the
dataset. This allows us to predict that a dataset size of about 7 million data points will be
required to achieve an accuracy of 99% in predicting the signal-to-noise ratio with the neural
network that we constructed.