Universal Gravitational Wave Parameter Estimation by Deep Learning
No Thumbnail Available
Date
2021
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
none
As the improvement of gravitational wave detectors, gravitationalwave events become more and more popular which opens a new win-dow of astronomy. In 2017, a binary neutron star event, GW170817,has been detected through the gravitational wave and also the electro-magnetic signal. After that, people start to consider an efficient wayto detect the GW and extract its dynamics parameters. In this thesis,we construct a Bayesian inference based on deep learning machine,CVAE, for the parameter estimation of binary black hole coalescence.This machine can obtain the inference of 5-dimensional parameters ofthe GW event within one second, where the parameters are two com-ponent mass m1 , m2 , luminosity distance dL , and time and phase ofcoalescence (tc , φ0 ). Since the noise of real detectors varies from timeto time, in contract to previous CVAE envelopments, we train ourmachine not only by strain data but also the corresponding amplitudespectrum density, which is used to characterize the noise background.We find our machine can obtain the compatible result in comparisonto traditional PE algorithm even with the noise drift, which meansthe noise background varies event by event. Finally, we apply ourmachine to the LIGO/Virgo third observing run (O3) events to testthe performance of our machine against real data.
As the improvement of gravitational wave detectors, gravitationalwave events become more and more popular which opens a new win-dow of astronomy. In 2017, a binary neutron star event, GW170817,has been detected through the gravitational wave and also the electro-magnetic signal. After that, people start to consider an efficient wayto detect the GW and extract its dynamics parameters. In this thesis,we construct a Bayesian inference based on deep learning machine,CVAE, for the parameter estimation of binary black hole coalescence.This machine can obtain the inference of 5-dimensional parameters ofthe GW event within one second, where the parameters are two com-ponent mass m1 , m2 , luminosity distance dL , and time and phase ofcoalescence (tc , φ0 ). Since the noise of real detectors varies from timeto time, in contract to previous CVAE envelopments, we train ourmachine not only by strain data but also the corresponding amplitudespectrum density, which is used to characterize the noise background.We find our machine can obtain the compatible result in comparisonto traditional PE algorithm even with the noise drift, which meansthe noise background varies event by event. Finally, we apply ourmachine to the LIGO/Virgo third observing run (O3) events to testthe performance of our machine against real data.
Description
Keywords
none, Gravitational wave, General relativity, Data analysis, Matched filter, Parameter estimation, Deep learning, Conditional variational autoencoder, Normalizing flow