Exxonmobil Technology and Engineering Company

Exxonmobil Upstream Research Company

Junzhe Sun

Mehdi Aharchaou

Erik R Neumann

Anatoly I Baumstein

Valeriy V Brytik

Rongrong Lu

Carey M Marcinkovich

Sirui Tan

Hong Zhao

Jia Yan

Harpreet Kaur

Lingxiao Zhang

A method and apparatus for generating a high-resolution seismic image, including extracting a reflectivity distribution from a geological model; utilizing the reflectivity distribution to label features of the model; generating forward-modeled data from the model; migrating the forward-modeled data to create a migrated image; and training a deep neural network with the labeled synthetic geological model and the migrated image to create a reflectivity prediction network. A method and apparatus includes: selecting a first subset of the field data; applying a low-pass filter to the first subset to generate a first filtered dataset; migrating the first filtered dataset to create a first migrated image; applying a high-pass filter to the first subset to generate a second filtered dataset; migrating the second filtered dataset to create a second migrated image; and training a deep neural network to predict a target distribution of high-frequency signal.

Method for generating initial models for least squares migration using deep neural networks

A methodology for extending bandwidth of geophysical data is disclosed. Geophysical data, obtained via a towed streamer, may have significant noise in a certain band (such as less than 4 Hz), rendering the data in the certain band unreliable. To remedy this, geophysical data, from a band that is reliable, may be extended to the certain band, resulting in bandwidth extension. One manner of bandwidth extension comprises using machine learning to generate a machine learning model. Specifically, because bandwidth may be viewed as a sequence, machine learning configured to identify sequences, such as recurrent neural networks, may be used to generate the machine learning model. In particular, machine learning may use a training dataset acquired via ocean bottom nodes in order to generate the machine learning model. After which, the machine learning model may be used to extend the bandwidth of a test dataset acquired via a towed streamer.

Bandwith extension of geophysical data

A method for approximating an inverse Hessian is provided. One methodology to generate the inverse Hessian is to precondition the gradient, such as by using point-spread function deconvolution, T-power, or source-illumination compensation, prior to using non-stationary matching filters (NMF) to generate the inverse Hessian. Various types of NMF are contemplated, including using filters for different windows in the subsurface or using filters assigned to specific locations in the subsurface. Further, the number of filters for NMF may vary from iteration to iteration. For example, the filters assigned to the specific locations in the subsurface may be generated in a multi-scale manner, in which an initial iteration uses longer scale/longer wavelength features for inversion and subsequent iterations use finer scale/smaller wavelength features for inversion.

Method for hydrocarbon prospecting using an approximated inverse hessian to update a property model

A method for partitioning a search direction when using least squares reverse time migration (LSRTM) is provided. LSRTM may be used iteratively in order to improve imaging accuracy. As part of LSRTM, multiple local line searches may be performed. In particular, image space may be partitioned, such as by using a set of masks. The search direction, such as the gradient, may be partitioned using the set of masks. Local line searches may be performed for each partition of the search direction, resulting in finding respective line search constants. The respective line search constants may then be used for iterating the model in order to improve imaging accuracy.

Method for partitioning a search direction when using least squares reverse time migration

A method for enhancing properties of geophysical data with deep learning networks. Geophysical data may be acquired by positioning a source of sound waves at a chosen shot location, and measuring back-scattered energy generated by the source using receivers placed at selected locations. For example, seismic data may be collected using towed streamer acquisition in order to derive subsurface properties or to form images of the subsurface. However, towed streamer data may be deficient in one or more properties (e.g., at low frequencies). To compensate for the deficiencies, another survey (such as an Ocean Bottom Nodes (OBN) survey) may be sparsely acquired in order to train a neural network. The trained neural network may then be used to compensate for the towed streamer deficient properties, such as by using the trained neural network to extend the towed streamer data to the low frequencies.

Methodology for enhancing properties of geophysical data with deep learning networks

A method and apparatus for generating attenuation-compensated images of subsurface region, including: computing an image of the region utilizing elastic wave propagation, based on field data and subsurface model; generating forward-modeled data utilizing forward viscoelastic wave propagation, based on the image; computing secondary image by migration; computing NMF based on the images; and applying the NMF to the image to generate the attenuation-compensated image. A method and apparatus includes: iteratively computing attenuation-compensated gradient of the region utilizing an elastic wave propagation operator in the back-propagation and a viscoelastic wave propagation operator in the forward modelling, based on field data and subsurface model; computing search direction based on the attenuation-compensated gradient, searching for an improved model, and checking the improved model for convergence.

Method for attenuation compensation utilizing non-stationary matching filters

A method, including: obtaining Earth models including velocity, anisotropy, and attenuation reconstructing a source wavefield using the Earth models; reconstructing a receiver wavefield using the Earth models, wherein the reconstructing the source wavefield and the receiver wavefield each include applying an attenuation operator that increases an amplitude of down-going wavefields within an attenuation body and that decreases an amplitude of up-going wavefields within the attenuation body; applying an imaging condition to the source wavefield and receiver wavefield for a plurality of shots; and generating a subsurface image by stacking images for the plurality of shots.

Growing community of inventors

San Jose, CA, United States of America

Junzhe Sun