Applications of supervised deep learning for seismic interpretation and inversion

Abstract

Recent advances in machine learning and its applications in various sectors are generating a new wave of experiments and solutions to solve geophysical problems in the oil and gas industry. We present two separate case studies in which supervised deep learning is used as an alternative to conventional techniques. The first case is an example of image classification applied to seismic interpretation. A convolutional neural network (CNN) is trained to pick faults automatically in 3D seismic volumes. Every sample in the input seismic image is classified as either a nonfault or fault with a certain dip and azimuth that are predicted simultaneously. The second case is an example of elastic model building — casting prestack seismic inversion as a machine learning regression problem. A CNN is trained to make predictions of 1D velocity and density profiles from input seismic records. In both case studies, we demonstrate that CNN models trained from synthetic data can be used to make efficient and effective predictions on field data. While results from the first example show that high-quality fault picks can be predicted from migrated seismic images, we find that it is more challenging in the prestack seismic inversion case where constraining the subsurface geologic variations and careful preconditioning of input seismic data are important for obtaining reasonably reliable results. This observation matches our experience using conventional workflows and methods, which also respond to improved signal to noise after migration and stack, and the inherent subsurface ambiguity makes unique parameter inversion difficult.