Paper submitted to SEG Asia Pacific Offshore Exploration Symposium (APEX), by Jyoti Kumar, Mark Roberts, Oystein Korsmo (TGS)
Summary
In this study, we present a pragmatic deep learning–based workflow designed to generalize across unseen field datasets and significantly reduce VMB turnaround time. The proposed approach targets three key components of the workflow: estimation of a robust initial macro-velocity model, automated picking of complex water-bottom horizons, and replacement of conventional tomography and interpretation steps to condition the initial FWI model. The deep learning approach, to obtain the initial FWI model, integrates Fourier Neural Operators (FNOs), convolutional neural networks (CNNs), and fully connected neural networks (FCNs) to estimate velocity errors in the image domain. We validate the approach using multiple field data examples from different sedimentary basins worldwide including the Agung area located north of Bali.