Abstract
Neural networks have been used for some time in geophysics to quantitatively predict elastic and rock properties from the seismic and well data. The traditional workflow usually involves inputting seismic attributes generated from the poststack seismic data and requires human input and expertise in their generation and selection. The optimal attributes change depending on the target so the workflow can be quite laborious. With the recent advent of deep neural networks (DNNs), it has become possible for the machine to learn the features automatically. This paper compares two DNN architectures to predict elastic and rock properties. The first architecture, a fully connected (FC) DNN uses as input handcrafted features. The second architecture, a convolutional neural network (CNN) uses as input seismic gathers segmented into a series of images. The generation and determination of the optimal features is implicitly performed as part of the training of the convolutional layers, avoiding the labour-intensive procedure of generating and selecting attributes. As this is done automatically, multiple targets can be estimated simultaneously providing a simpler and more efficient workflow.
In these supervised learning approaches, the seismic-to-rock property relationship is learned from the data. One of the major factors limiting the success of these methods is whether there exists enough labelled data to train the neural network adequately. To overcome these issues, we generate synthetic data. First, we simulate many pseudo-wells based on the well statistics in the project area. In addition, the reservoir properties, such as porosity, saturation, mineralogy and thickness, are all varied to create a well-sampled dataset. Elastic and synthetic seismic data are then generated using rock physics and seismic theory. The resulting collection of pseudo-well logs and synthetic seismic data, called the synthetic catalog, is used to train the neural network. The derived operator is then applied to the real seismic data to predict reservoir properties throughout the seismic volume.
This synthetic data workflow is applied to a Gulf Coast dataset. In this study only one well is used to generate hundreds of synthetic wells and seismic gathers. Any of the simulated logs can serve as the target. In this case, the P-wave impedance, P-wave to S-wave velocity ratio, density and saturation are estimated. The synthetic wells and seismic gathers are used to train both the FC DNN and CNN. The input to the FC DNN are seismic attributes generated from angle stacks optimally determined for each target. Hence, the FC DNN must be trained on each target, in this case four times. In contrast, the input to the CNN are the seismic gathers. The CNN is trained simultaneously for all four targets, providing a more efficient workflow. A key to successfully training both these networks is the big data supplied by the synthetic catalog.
Biography
Jon Downton is a Senior Research Advisor with CGG GeoSoftware focused on the HampsonRussell Software Development. Over his 35 years in the industry, Jon has worked as a reservoir geophysicist, research geophysicist and research manager. He was a cofounder of Integra Geoservices. After selling Integra to Core Labs in 1999, he went back to university and did his Ph.D. on AVO inversion. Since that time, his research has focused on reservoir geophysics and the seismic processing associated with this. His current research focus is the application of machine learning to the field of reservoir geophysics. Jon has extensive experience in estimating rock and fluid properties from seismic data, including AVO, AVAz inversion and rock physics. Jon is a member of the CSEG, SEG, EAGE and APEGA and is a past president of the CSEG.
Dan Hampson received a BSc degree in physics from Loyola College, Montreal, in 1971, and received an MSc in theoretical physics from McMaster University, Hamilton, in 1973. In 1976, Dan joined Veritas Seismic Processors, where he held several positions, including Vice President for Research. In 1987, Dan left Veritas and joined with Brian Russell to form Hampson-Russell Software Services, where he was president from 1987 to 2002. In 1996, Dan and Brian received the SEG Cecil Green Enterprise Award, and in 2006, Dan received the SEG Reginald Fessenden Award for work on the Parabolic Radon Transform. In 2002, Hampson-Russell was acquired by VeritasDGC, subsequently CGG, where Dan continues in a part-time advisory role, collaborating with Jon and Brian on research to apply Machine Learning to reservoir characterization.
