Artificial Intelligence Improves Seismic-Image Reconstruction

Abstract

This article, written by JPT Technology Editor Chris Carpenter, contains highlights of the open-submission paper “Artificial Intelligence for Seismic-Image Reconstruction,” by Yogendra Narayan Pandey, SPE, and Govind Chada, Prabuddha, and Tejas Karmarkar, Oracle Cloud Infrastructure. The paper was not presented at an SPE conference and has not been peer-reviewed. Seismic imaging provides vital tools•for the exploration of potential hydrocarbon reserves and subsequent production-planning activities. The acquisition of high-resolution, regularly sampled seismic data may be hindered by physical or financial constraints, which lead to undersampled, sparse seismic data. However, if seismic data are available at a higher resolution and sampled evenly throughout the region of interest, the generated 3D models of petrophysical properties could be improved. Such improvements would show potential benefits through the successive steps of reservoir modeling and production•planning. Traditional Approaches Traditional methods used to overcome the previously mentioned data-quality issues can be divided broadly into three categories. Wave-equation-based methods. These methods use physics-based•wave-propagation equations, using•velocity models•to•reconstruct missing seismic traces. Domain-transform methods. These are data-driven methods that involve transformation of data•between different domains, such as time and frequency. Prediction-error filters. These methods use a filter that learns from the known seismic data and constructs missing seismic data. Given recent advances in the field of artificial intelligence (AI), it is worth examining whether AI methods also can be useful in the task of seismic-data•reconstruction. Generative Adversarial Networks (GANs) GANs are a recent addition to the field of solution techniques. A variant of GANs, conditional generative adversarial networks, was used to test the efficacy of GANs for seismic-data reconstruction. Fig. 1 shows a schematic of a GAN model for seismic-data reconstruction. To effectively reconstruct a seismic image, a large number of 2D seismic images are fed into the GAN during the training process. A GAN model consists of two main components: Generator. This is a deep convolutional neural network (CNN) that uses random noise as an input and generates an image by expanding the input through a series of deconvolutions. As shown in Fig.•1, the generator network is fed 2D seismic images, with a portion of these masked to replicate missing traces. The generator network tries to reconstruct these•masked portions of the 2D seismic images. Discriminator. This is also a deep CNN that is shown the images reconstructed by the generator network and the original seismic images from which some parts were masked. The discriminator’s job is to distinguish between the images reconstructed by the generator from the original seismic images. As training progresses, the generator network tries to create images that the discriminator will not be able to differentiate from the original seismic images. At the same time, the discriminator’s ability to distinguish the generated images from the original images continuously improves. As a result, the generator begins producing reconstructed seismic images, which look similar to the original ones.