Gravitational wave surrogate modeling
Deep Learning methods in surrogate modeling of gravitational waveforms.
Overview
Gravitational waves are produced when massive astrophysical objects merge in a violent manner. The usual participants in these events are neutron stars and black holes, both stellar corpses, remnants of stars that are long gone. When these objects coalesce, the space around them vibrates and ripples are produced (imagine the surface of a lake after the landing of a leaf) and because they interact with matter faintly, they can travel very very (very!) long distances. Gravitational waves were predicted by Albert Einstein back in 1916 but were first observed almost 100 years later in September 14, 2015 by the twin Laser Interferometer Gravitational-wave Observatory (LIGO) detectors!
One field that flourished was that of gravitational wave modeling, where one must solve the Einstein equations either via analytical or numerical solutions… But, because these simulations are computational heavy, scientists turned to the approximation of these simulated gravitational waveforms via surrogate models which were much lighter but quite accurate! However, as with most methodologies, surrogate modeling had its own issues and limitations. So people working in this field, tunred to a -at that time- rapidly spreading tool: artificial neural networks, to replace one of the steps of the generation of surrogate waveforms.
In the project, in which I participated as a student, we were looking for patterns in the data and computational methodologies to improve these models.
A collaboration project between the CIDL research group and the Gravitational Waves Group, AUTh from the departments of Informatics and Physics of Aristotle University of Thessaloniki.
My contribution focused on exploring deep learning architectures as well as applying various tricks to reduce the learning errors.
Related Publication
Results from this project were published in Elsevier’s Applied Soft Computing and Neurocomputing.