Search

EP-3818398-B1 - CASCADED MACHINE-LEARNING WORKFLOW FOR SALT SEISMIC INTERPRETATION

EP3818398B1EP 3818398 B1EP3818398 B1EP 3818398B1EP-3818398-B1

Inventors

  • KAUL, Anisha
  • LI, CEN
  • MANIAR, Hiren
  • ABUBAKAR, ARIA

Dates

Publication Date
20260513
Application Date
20190626

Claims (10)

  1. A computer-implemented method, comprising: (a) receiving raw seismic data by a computer system, the seismic data being acquired in a seismic survey and being in the form of a three dimensional seismic volume having an inline direction, a crossline direction and a depth direction and comprising a plurality of inline slices, each comprising a two-dimensional vertical surface of the seismic volume having a fixed inline coordinate value, a plurality of crossline slices, each comprising a two-dimensional vertical surface of the seismic volume having a fixed crossline coordinate value and a plurality of depth slices, comprising a two-dimensional horizontal surface of the seismic volume having a fixed depth coordinate value; (b) training a first model using the plurality of crossline and inline slices, wherein the first model is a supervised machine learning model of a top of salt TOS surface in the seismic volume; (c) predicting, by the machine-learning procedure of the computer system, cubes of seismic data in the inline and crossline directions based on the model and combining the predicted cubes; (d) generating, by the machine-learning procedure of the computer system, a probability cube of TOS labels based on the combined predicted cubes and applying threshold to the probability cube to generate a binary cube where 1 = salt and 0 = no salt; (e) determining, by the machine-learning procedure of the computer system, a depth D in the seismic volume above which a predicted TOS surface has a predetermined level of clarity; (f) sampling (912), by the machine-learning procedure of the computer system, seismic data in the depth direction from a top of the seismic volume to the determined depth D to obtain a two-dimensional training seismic slice; (g) sampling (916) by the machine-learning procedure of the computer system, the binary cube in the depth direction above the determined depth D to obtain a two dimensional mask slice, the mask slice comprising a slice of the binary cube; (h) selecting (918), by the machine-learning procedure of the computer system, a first pixel in the training seismic slice and a second pixel in the mask slice, wherein the first and second pixels have common inline and crossline coordinate values; (i) performing (922), by the machine-learning procedure of the computer system, one or more recursions of step (h) for first and second pixels of the training slice and mask slice to generate a plurality of pairs of pixel coordinate values; (j) generating or updating (928), by the machine-learning procedure of the computer system, a model of the seismic volume based upon the generated pairs of pixel coordinate values; and (k) determining (1018), by the machine-learning procedure of the computer system, an amount or location of hydrocarbons in the seismic volume using the generated or updated model.
  2. The method of claim 1, further comprising: (l) sampling (1006), by the machine-learning procedure of the computer system, the seismic data in the seismic volume to obtain a two-dimensional evaluation seismic slice, wherein the seismic data is sampled from the top of the seismic volume to a bottom of the seismic volume to obtain the evaluation seismic slice and wherein the depth is between the top and the bottom.
  3. The method of claim 2, further comprising: (m) determining (1008), by the machine-learning procedure of the computer system, a presence of a salt body in the evaluation seismic slice based upon the seismic data in the evaluation seismic slice.
  4. The method of claim 3, further comprising: (n) generating or updating (1010), by the machine-learning procedure of the computer system, a three-dimensional (3D) matrix based upon the model and the evaluation seismic slice, wherein the 3D matrix indicates the presence of the salt body.
  5. The method of claim 4, further comprising: extracting (1014), by the machine-learning procedure of the computer system, a geobody from the 3D matrix; and extracting (1016), by the machine-learning procedure of the computer system, a surface from the geobody.
  6. The method of claim 5, further comprising determining, (1018) by the machine-learning procedure of the computer system, a presence of hydrocarbons in the seismic volume based upon the 3D matrix, the geobody, the surface, or a combination thereof.
  7. The method of claim 1, further comprising performing, by the machine-learning procedure of the computer system, a plurality of recursions of steps (e) to (i) for a plurality of depth values.
  8. The method of claim 4, wherein the 3D matrix has a same size and shape as the seismic volume.
  9. The method of claim 2, 3 and 4, further comprising performing (1012), by the machine-learning procedure of the computer system, a plurality of recursions of steps (1), (m) and (b) on a plurality of evaluation seismic slices.
  10. A computing system (1200) comprising: one or more processors (1204); and a memory system (1206) comprising one or more non-transitory computer-readable media storing instructions that, when executed by at least one of the one or more processors (1204), cause the computing system (1200) to perform the steps of the method of any preceding claim.

Description

Background A robust earth model may be used to create a high-fidelity image of a subterranean formation. Often, earth models include geological features such as salt (e.g., halite) bodies and seismic facies. Delineating or interpreting these geological features may be used to model salt velocity. In addition, seismic interpretation may be used to generate an accurate and geologically-sound annotation of an oil and gas survey area (e.g., a well site above the subterranean formation). Examples of interpretation activities include geobody extraction, fault interpretation, horizon interpretation, and salt interpretation. Multiple interpreters may spend months interpreting horizons to isolate salt bodies with the goal of creating an accurate structural earth model encompassing thousands of squared kilometers. This involves manually selecting horizons on a coarse grid. This selection process may be iterated until a surface meets predetermined criteria within the subterranean formation. The underlying effort may vary with both the complexity of the salt boundaries, as well as the size of the seismic survey. Thus, as will be appreciated, this process may be both laborious and highly biased by an interpreter's view and experience. Nevertheless, salt diapirs are useful geological features because such formations act as natural traps for hydrocarbons (e.g., oil and gas). Amin Asjad et al, "Salt-Dome Detection Using a Codebook-Based Learning Model", IEEE GEOSCIENCE AND REMOTE SENSING LETTERS, IEEE, USA, vol. 13, no. 11, 1 November 2016, pages 1636 - 1640, describes a learning model for salt-dome detection in seismic imaging using texture-based attributes. The algorithm works by combining the attributes from the gray-level cooccurrence matrix (GLCM) and those from the Gabor filter, with a codebook-based learning approach to delineate salt boundaries in seismic data. US 2016/086352 describes a method for detecting a mineral layer in seismic survey image data including transforming the intensity of an unprocessed seismic survey image volume in the form of a 3-dimensional (3D) grid of voxels each associated with an intensity so as to enhance a contrast of the seismic survey image volume. The intensity transformed image is scanned voxel-by-voxel with a classifier to determine a probability of each voxel being associated with a mineral layer, Thresholds are applied to the voxel probabilities to yield a 3D binary image mask that corresponds to the seismic survey image volume, wherein each voxel of the binary image mask has a value indicative of whether the voxel is mineral or non-mineral. Ferreira Rodrigo da Silva et al, "Multi-scale Evaluation of Texture Features for Salt Dome Detection", 2016 IEEE INTERNATIONAL SYMPOSIUM ON MULTIMEDIA (ISM), IEEE, 11 December 2016, pages 632 - 635 describes detecting salt domes by performing a comprehensive evaluation of texture descriptors broadly used in the image processing community when applied to seismic images. A robust multi-scale analysis is conducted in order to assess which features and corresponding parameters are more relevant for salt dome detection according to various patch sizes. Summary The present invention resides in a computer-implemented method as defined in claim 1 and in a computing system as defined in claim 10. Brief Description of the Drawings The accompanying drawings, which are incorporated in and constitute a part of this specification, illustrate embodiments of the present teachings and together with the description, serve to explain the principles of the present teachings. In the figures: Figures 1A, 1B, 1C, 1D, 2, 3A, and 3B illustrate simplified, schematic views of an oilfield and its operation, according to an embodiment.Figure 4A illustrates a flowchart of a crossline/inline-based machine-learning training workflow, according to an embodiment.Figure 4B illustrates a flowchart of a crossline/inline-based machine-learning prediction workflow, according to an embodiment.Figure 5A illustrates a model trained by a seismic interpreter, according to an embodiment.Figure 5B illustrates a model trained on crossline and inline directions of a seismic survey where the flanks are poorly-captured and the resulting predicted surfaces are discontinuous, according to an embodiment.Figure 6 illustrates a schematic view of a machine-learning-based procedure to improve salt-body detection accuracy and reduce false negatives, according to an embodiment.Figure 7A illustrates an image including dense labels in the crossline and inline directions that lead to a dense salt body when viewed along the depth slice, according to an embodiment.Figure 7B illustrates an image including sparse labels across the crossline and inline directions, according to an embodiment.Figure 8A illustrates an image showing the amplitude distribution of the crossline direction, according to an embodiment.Figure 8B illustrates an image showing the amplitude distribution of the inline direction, accordin