Deep learning in seismic wavefield reconstruction and resolution enhancement

Seismic data contain information about elastic properties and the structural patterns of the subsurface. However, before we get an image of the subsurface, seismic data goes through many processing steps that consist of applying theory of wave propagation and mathematical and statistical signal processing. In this thesis, we have investigated breakthroughs within artificial intelligence for seismic wavefield reconstruction. In the first paper, we developed a method that is suited for marine collection where one can utilize similarities between acquisition domains to train the machine learning model. In the second paper, we utilize methodology from image inpainting combined with inverse theory to fit the reconstruction problem in the offset domain, within an unsupervised deep learning framework. In the final paper, we illustrate how one can demigrate field data to generate suitable training data for interpolation and regularization of offset classes. This synthesized training data can then be used to create the prediction models. In all three papers, we have documented an improvement compared to existing "state-of-the-art" methods from the industry.