We propose a workflow to obtain better shear-wave velocity initial models for elastic full-waveform inversion (EFWI). The workflow consists of a convolutional neural network (CNN) to estimate the shear-wave velocity (Vs) model from a seismic image, constrained by the compressional-wave velocity (Vp) model, given the well logs. More specifically, the workflow starts with the seismic shot gathers and an initial Vp model (e.g., from the traveltime tomography) to generate an image. The CNN is trained with the seismic image as an input, well logs as output properties, and the initial Vp model as a constraint. After training, the network estimates a Vs model, which is used as an initial model for the first frequency band of EFWI, to update the Vp model. This process is repeated with an updated Vp model and an updated migrated seismic image to generate an updated initial Vs model for the next frequency band of EFWI. We demonstrate the effectiveness of the proposed EFWI workflow on the Marmousi 2 synthetic dataset, with the estimated initial Vs models from reverse time migration images and well logs by a CNN. Improvements on the continuity and resolution of the estimated Vp model are observed when using the data-driven estimated initial Vs models, compared to using the Vs models from a fixed ratio between Vp and Vs in well logs.

Time-lapse (4D) seismic is an essential and reliable tool for monitoring geologically sequestered CO 2 in the subsurface. Unfortunately, 4D seismic is expensive due to repeated data acquisition and the complex procedures involved in 4D seismic data processing. To reduce the cost of time-lapse monitoring of injected CO 2 plume bodies, we have proposed three novel strategies powered by deep learning. Each of these strategies is driven by customized deep learning networks and specialized training schemes to meet both the requirements of monitoring accuracy and the constraints of data availability. These strategies address challenging problems at various stages of time-lapse seismic monitoring, ranging from seismic data acquisition to data processing. Specifically, the first strategy was designed to reconstruct sparsely acquired seismic data into a complete dense data set without losing critical subsurface information. The resulting reconstructed dense data set can then be fed into conventional seismic processing modules without suffering from the aliasing issue. Consequently, the first strategy does not reduce the cost of 4D data processing. The second strategy attempts to bypass time-consuming conventional physics-based seismic data processing procedures by leveraging the powerful nonlinear mapping capabilities of neural networks, which allows us to use sparser data for CO 2 monitoring in situations where data reconstruction is unfeasible due to a lack of coherence in these sparser data sets. Therefore, this strategy reduces data processing cost but is more sensitive to noise. The third strategy is a unique workflow to further cut monitoring costs by replacing 3D seismic data with multiple 2D lines combined with optional point seismic measurements to reconstruct full 3D seismic image volumes or 3D subsurface property models for economical CO 2 monitoring. While this strategy features the most cost-effective data acquisition schemes, it is important to note that high-wavenumber structural information between sparse 2D lines cannot be fully recovered.

Time-lapse seismic is one of the most effective tools for monitoring subsurface processes or property changes caused by hydrocarbon production, CO2 sequestration, geothermal development, and many other activities. To substantially reduce the cost and turnaround time of time-lapse seismic data processing projects, we developed a deep-learning network for rapid characterization of subsurface property changes by establishing direct nonlinear mapping from premigration seismic data to subsurface property models. By bypassing the time-consuming conventional seismic processing procedures, such as seismic data migration and full waveform inversion (FWI), this deep-learning network enables us to quantitatively estimate the subsurface condition changes almost instantaneously by efficiently scanning, selecting, and analyzing any new monitoring datasets. This property estimation network architecture features a multibranch design with different convolutional filtering sizes for better feature extraction from dipping events within the seismic gathers. In addition, a customized loss function with a weighted term is used to address the imbalanced training label issue. Furthermore, to effectively suppress the time-lapse data nonrepeatability for accurate time-lapse response extraction from the seismic data, we developed an additional deep-learning network known as the repeatability enforcement (RE) network. This network features a specially designed learning strategy aimed at eliminating differences between baseline and monitoring data induced by non-CO2-related factors, such as variations in bandwidth, source signature, and seasonal changes in near-surface conditions, etc. By combining the property estimation network with the RE network, we demonstrate the successful application of this deep-learning approach to both the Sleipner data and Chimera data.

PreviousNext No AccessGEOPHYSICSJust-Accepted ArticlesLatest advancements in machine learning for geophysics — IntroductionAuthors: Haibin DiWenyi HuAria AbubakarPandu DevarakotaWeichang LiYaoguo LiHaibin Di SLB, Digital Subsurface Intelligence, Houston, Texas, USA. E-mail: [email protected]; [email protected]; [email protected].Search for more papers by this authorEmail the author at [email protected]Email the author at [email protected]Email the author at [email protected], Wenyi Hu SLB, Digital Subsurface Intelligence, Houston, Texas, USA. E-mail: [email protected]; [email protected]; [email protected].Search for more papers by this authorEmail the author at [email protected]Email the author at [email protected]Email the author at [email protected], Aria Abubakar SLB, Digital Subsurface Intelligence, Houston, Texas, USA. E-mail: [email protected]; [email protected]; [email protected].Search for more papers by this authorEmail the author at [email protected]Email the author at [email protected]Email the author at [email protected], Pandu Devarakota Shell Technology Center Houston, Deep Learning and AI, Houston, Texas, USA. E-mail: [email protected].Search for more papers by this authorEmail the author at [email protected], Weichang Li Aramco Americas, Houston Research Center, Houston, Texas, USA. E-mail: [email protected].Search for more papers by this authorEmail the author at [email protected], and Yaoguo Li Colorado School of Mines, Department of Geophysics, Golden, Colorado, USA. E-mail: [email protected].Search for more papers by this authorEmail the author at [email protected]https://doi.org/10.1190/geo2023-1116-spseintro.1 SectionsAboutPDF/ePub ToolsAdd to favoritesDownload CitationsTrack Citations ShareFacebookTwitterLinked InReddit FiguresReferencesRelatedDetails Just-Accepted ArticlesPages: 1-96ISSN (print):0016-8033 ISSN (online):1942-2156 publication data© 2024 Society of Exploration GeophysicistsPublisher:Society of Exploration Geophysicists HistoryPublished Online: 20 Nov 2023 CITATION INFORMATION HaibinDi, WenyiHu, AriaAbubakar, PanduDevarakota, WeichangLi, and YaoguoLi, (), "Latest advancements in machine learning for geophysics — Introduction," GEOPHYSICS 0: 1-4. https://doi.org/10.1190/geo2023-1116-spseintro.1 Plain-Language Summary PDF Download Metrics Loading ...

We develop two unique deep learning algorithms to tackle the repeatability issues of time-lapse seismic measurements. The supervised approach, based on the image-to-image transformation algorithm, includes a customized training schematic based on the plume body extension. Meanwhile, the unsupervised method, based on the mechanism of image registration, automatically performs the event alignments to minimize event misfits of the time-lapse seismic images. We apply both methods to a field dataset for performance evaluation, comparing them to an existing conventional time-lapse processing workflow.

Seismic inversion is the primary approach for converting seismic images into geophysical property models for subsurface interpretation and reservoir characterization; however, the traditional workflow is complicated and requires massive computational resources and human supervision. Here we present a practical workflow that significantly automates the process, which consists of four major steps, (i) large-scale structural model construction, (ii) initial property model estimation via a multi-task convolutional neural network (CNN), (iii) physics-based property-reflectivity-seismic pair generation, and (iv) property model estimation refinement via a physics-guided CNN. Its application to the public Volve dataset in North Sea successfully produces density and P-slowness volumes of high lateral consistency and vertical resolution. Synthetic seismic is also derived and observed in close correlation with the actual seismic images, which further validate the accuracy of machine prediction by the proposed workflow.

We introduce a novel three-stage machine learning (ML) workflow for seismic resolution enhancement by integrating well log measurements, specifically designed for dataset with limited number of wells. The algorithm includes a pseudo well-log generator from input seismic volumes. The resolution enhancement is then performed by mapping from the low- to high-resolution seismograms of many randomly chosen pseudo well logs from the generator. Comparing to the existing supervised learning approaches that suffer from the network generalization challenge posed by limited number of available wells, the proposed workflow greatly increases the reliability of resolution-enhanced image for sparse well scenarios, as it exploits the realistic local geological information carried in the large number of pseudo-wells during the network training.

In this research, we combine data-driven and physics-driven strategies to train a physics-informed neural network for the simulation of electromagnetic wave propagation. By leveraging the constraints from the partial differential equation, the pre-trained network is able to efficiently predict the electromagnetic field in the entire computational domain with a small amount of training data, yielding more accurate results than the pure data-driven neural networks. Despite the recognized challenges, the proposed method shows a potential to accelerate and stabilize the inversion process when the amount of measurement data is limited due to data acquisition constraints under certain circumstances.

The simultaneous-source technology for high-density seismic acquisition is a key solution to efficient seismic surveying. It is a cost-effective method when blended subsurface responses are recorded within a short time interval using multiple seismic sources. A following deblending process, however, is needed to separate signals contributed by individual sources. Recent advances in deep learning and its data-driven approach toward feature engineering have led to many new applications for a variety of seismic processing problems. It is still a challenge, though, to collect enough labeled data and avoid model overfitting and poor generalization performance over different datasets with a low resemblance from each other. In this article, we propose a novel self-supervised learning method to solve the deblending problem without labeled training datasets. Using a blind-trace deep neural network and a carefully crafted blending loss function, we demonstrate that the individual source-response pairs can be accurately separated under three different blended-acquisition designs.

The evolution of digitization and advancements in seismic data acquisition technology have led to an exponential increase in data volumes within the seismic industry. This surge in data has introduced significant challenges in storage, transmission, and computation. Despite the attempts of previous studies to explore seismic data compression, achieving a balance among the compression ratio, quality of decompressed data, and computational efficiency continues to pose a challenge. This study capitalizes on the recent advancements in deep learning to utilize a variational autoencoder incorporated with a hyperprior. The aim is to simultaneously optimize the reconstruction accuracy and compression ratio in a comprehensive, end-to-end approach. Our experimental analyses demonstrate the efficacy of the introduced compression model on both pre-migration and post-migration seismic datasets.

The full waveform inversion (FWI) is highly non-linear and underdetermined. Due to the limitation of acquisition devices, low-frequency information may be missing from obtained data. The absence of low-frequency information could severely compromise FWI results. To solve these problems, we propose a learning-based approach to extrapolate the low-frequency data using a progressive transfer learning workflow to eliminate the labeling requirement for training. The system is implemented on a differentiable program platform to take advantage of automatic differentiation. The performance of the system is validated by an electromagnetic (EM) inverse problem. Our experiment results show that the proposed method is efficient and effective.