Automatic fault interpretation from seismic data via convolutional neural networks

Abstract

Summary These days understanding of fault geometry distribution across a particular oil or gas reservoir becomes very important task. It arises from the fact that fluid flow of present unconventional deposits is mostly driven by natural fractures instead of sedimentary porosity and corresponding permeability. On the other, complex compartmentalized reservoir separated into small discontinuous deposits by tectonic activity could lead to economic risks during field development. Most of conventional tools for fault interpretation from seismic data are highly affected by noise from data and deterministic so cannot produce probabilistic output. In the presented research application of convolutional neural networks for fault interpretation from seismic data was considered. Proposed architecture and training process were described. It was shown by metrics and visual analysis that developed method is able to delineate faults from seismic data in different geological and geophysical conditions. Additional advantage of suggested approach is its ability to produce probabilistic output allowing robust work with geological uncertainties and economic risks related to them due to consideration of many probable cases.