A machine learning approach to the search for gravitational waves emitted by light systems

With GW170817, gravitational waves have shown themselves to be very useful for multi-messenger astronomy. Combining the information from multiple channels such as gravitational waves, gamma-rays, neutrinos, etc. can lead to great physics. Contrarily to the electromagnetic telescopes, a gravitational wave interferometer surveys the entire sky. They do not have to focus on a small portion of the celestial sphere as do standard telescopes. It is also known that for binary neutron stars, the electromagnetic counterpart is produced during the last phase of the merger, whereas the gravitational wave signal can be detected several minutes before these last stages. If one is able to detect this signal before the merger and infer the sky location, gravitational wave astronomy can then send an alert and produce a sky map indicating where the astronomer can point their telescopes to see an electromagnetic counterpart.

The standard technique to detect these compact binary coalescences is matched filtering. The principle is to compute a template bank of pre-computed waveforms and match them with the data strain coming from the LIGO and Virgo interferometers. This thesis starts by illustrating a matched filter search with a project to detect long signals coming from sub-solar coalescence.

Recently, some matched filtering pipelines have started to adapt their method to search for gravitational waves with only the early stage of the signal. Other methods are beginning to be developed for this type of research. This thesis presents new methods, based on machine learning, to detect the early phase of a binary neutron star merger. We have developed multiple convolutional neural networks looking directly at the strain data of the detector to detect binary neutron stars before the merger.

The last step to produce an early warning for the astronomer is to create a sky map indicating the location of the event. We therefore shortly discuss how to accomplish this through a machine learning method for the whole signal, and also mention how it can be adapted to the early part of the signal.