The detection of gravitational waves from the coalescence of compact binaries formed by black holes and/or neutron stars strongly relies on accurate waveform templates to conduct searches based on matched filtering. Faithful templates can be built either by solving Einstein's gravitational-field equations with numerical-relativity techniques or by using approximations to the two-body problem in general relativity. This is a computationally demanding task since millions of templates must be generated in order to have a good accuracy not only for the confirmation of a detection event, but also for the characterization of the event. Our proposal offers a new approach to tackle this problem by using self-supervising learning to build deep-learning models capable of extracting features from the signals by using the combination of available LIGO-Virgo public data and a few simulated waveforms.
Artemisa has been instrumental to achieve our goal of performing gravitational-wave parameter inference using Bayesian deep learning, enabling us to generate significantly larger datasets and use deeper/more sophisticated network architectures. Furthermore, we have also used Artemisa to develop quick scan scripts for probing new physics using primordial gravitational waves and to generate binary neutron star merger waveforms using Generative Adversarial Networks. Our proposal also makes use of deep learning to study the burst gravitational-wave signals produced in supernova explosions of massive stars. To this goal we train a network for signal classification and separation from transient sources of detector noise (glitches) and a second network for inference of physical parameters of the new-born neutron stars. Ultimately, our program intends that the methods we are testing in Artemisa contribute to the toolset of the Virgo Collaboration.
