Deep learning is arguably one of the most important innovations in artificial intelligence in recent times. It allows for computational solutions to problems that are not easily characterized by a mathematical model or deterministic algorithm. It also allows for automated solutions to problems that are inherently subjective. Both of these criteria are endemic in the earth sciences, so innovative solutions to these challenges should be welcomed. We demonstrate a recent refinement to a deep-learning fault identification process that improves the continuity and compactness of predicted fault planes in areas where faults intersect. Historically, predictions from both deep learning and traditional algorithmic approaches were characterized by “blurry” clouds of intermediate probability values that extended well beyond the fault plane. To remediate this blurring problem and enhance confidence of inferences, we demonstrate a preprocessing technique in the image domain by using generative adversarial networks (GANs) that sharpen the seismic image prior to training and prediction. This sharpening solution consists of two neural networks. A feature-extraction network is used for extracting both local and global features from an unrelated, high-quality “donor” seismic survey. Then, the data set of interest is sent through a donor reconstruction network where a generator architecture creates plausible-looking images at a denser sampling rate with high perceptual quality. In the study presented here, we create our sharpening network using a modern, high-fidelity, deepwater 3D survey with well-imaged faults as the donor. The resulting generator architecture is then applied to our data set of interest — an altogether separate deepwater data set in the Gulf of Mexico. Similar in intent to a 5D interpolation, the GANs-based supersampled data contain three times the inline and crossline trace density, and the sampling interval is upsampled by a factor of three. This approach aims to preserve spatial and temporal frequency content of the parent data while providing a denser data set for deep-learning applications. The supersampled data is then deployed into our deep-learning training regimen to enhance the performance of our fault detection network. By introducing a preprocessing sharpening step, the predicted faults are less blurry, more compact, and more amenable to programmatic attempts to segment them into discrete features.

You do not currently have access to this article.