Granted Patent B2

US 12,399,291 · App. 17/595,002 · Granted Aug 26, 2025

Inventors: Vishakh Hegde (Menlo Park, CA); Suhas Suresha (Menlo Park, CA); Carlos Boneti (Saratoga, CA); Sergey Doronichev (Menlo Park, CA); Jose Celaya Galvan (Menlo Park, CA); Anthony Lichnewsky (Gatwick, GB)

Assignee: SCHLUMBERGER TECHNOLOGY CORPORATION

G01V1/302G01V1/306G01V1/345G01V20/00G06N3/04G06N3/08G01V2210/66

Abstract

Claims (46)

1. A subsurface structure identification system, comprising:

one or more processors; and

a memory coupled to the one or more processors, and encoded with instructions that when executed by the one or more processors cause the one or more processors to:

provide a convolutional neural network trained to identify subsurface structures in input migrated seismic volumes;

partition a first input migrated seismic volume into three-dimensional sub-volumes, wherein the three-dimensional sub-volumes are overlapping portions of the first input migrated seismic volume;

process each of the three-dimensional sub-volumes in the convolutional neural network to produce a respective plurality of feature maps associated with each three-dimensional sub-volume;

generate, for each three-dimensional sub-volume, a respective segmentation that identifies portions of the three-dimensional sub-volume as either belonging to sub-surface structure or background based on the respective plurality of feature maps associated with the three-dimensional sub-volume; and

identify a first subsurface structure in the first input migrated seismic volume based on a probability map of the first input migrated seismic volume generated based on the segmentations associated with the three-dimensional sub-volumes.

2. The subsurface structure identification system of claim 1 , wherein the convolutional neural network comprises a convolutional layer configured to provide volumetric convolution of seismic data.

3. The subsurface structure identification system of claim 1 , wherein the instructions, when executed by the one or more processors, further cause the one or more processors to normalize amplitude of the first input migrated seismic volume.

4. The subsurface structure identification system of claim 1 , wherein the instructions, when executed by the one or more processors, further cause the one or more processors to statistically aggregate a plurality of probability values produced by processing the three-dimensional sub-volumes.

5. The subsurface structure identification system of claim 1 , wherein the instructions, when executed by the one or more processors, further cause the one or more processors to:

apply a decision function to a plurality of probability values produced by processing the three-dimensional sub-volumes; and

determine a pixel state based on a result of the decision function for each of the plurality of probability values.

6. The subsurface structure identification system of claim 1 , wherein the first subsurface structure is a channel.

7. The subsurface structure identification system of claim 1 , wherein a first three-dimensional sub-volume overlaps a second three-dimensional sub-volume by no more than 30%.

8. The subsurface structure identification system of claim 1 , wherein the respective plurality of feature maps associated with each three-dimensional sub-volume is respective two feature maps associated with each three-dimensional sub-volume.

9. The subsurface structure identification system of claim 1 , wherein the portions of each three-dimensional sub-volume that are identified as either belonging to sub-surface structure or background are represented by pixels.

10. A method for identifying a subsurface structure, comprising:

applying a plurality of multi-dimensional sub-volumes of a first input migrated seismic volume to train a volumetric convolutional neural network to identify a subsurface structure in the first input migrated seismic volume, and produce a trained convolutional neural network;

partitioning a second input migrated seismic volume into multi-dimensional sub-volumes of seismic data, wherein the multi-dimensional sub-volumes are overlapping portions of the second input migrated seismic volume;

processing each of the multi-dimensional sub-volumes of the second input migrated seismic volume by the trained convolutional neural network to produce a respective plurality of feature maps associated with each multi-dimensional sub-volume of the second input migrated seismic volume;

generating, for each sub-volume of the second input migrated seismic volume, a respective segmentation that identifies portions of the multi-dimensional sub-volume as either belonging to sub-surface structure or background based on the respective plurality of feature maps associated with the multi-dimensional sub-volume;

generating a probability map for the second input migrated seismic volume based on the segmentations associated with the multi-dimensional sub-volumes of the second input migrated seismic volume; and

identifying the subsurface structure in the second input migrated seismic volume based on the probability map;

wherein the multi-dimensional sub-volumes are at least three dimensional.

11. The method of claim 10 , further comprising applying biased sampling to select the multi-dimensional sub-volumes of the first input migrated seismic volume that are applied to train the volumetric convolutional neural network.

12. The method of claim 11 , wherein applying the biased sampling comprises:

selecting a first pixel of the first input migrated seismic volume, wherein the first pixel has at least a predetermined probability of being part of the subsurface structure; and

selecting a set of pixels about the first pixel to form one of the multi-dimensional sub-volumes of the first input migrated seismic volume.

13. The method of claim 12 , further comprising varying the predetermined probability in different instances of training the volumetric convolutional neural network.

14. The method of claim 10 , wherein training the volumetric convolutional neural network comprises:

applying one or more back-propagation based optimization loops to minimize a weighted loss function that is applied to the volumetric convolutional neural network; and

deriving values of weights of the weighted loss function from the first input migrated seismic volume.

15. The method of claim 10 , further comprising normalizing amplitude of the second input migrated seismic volume prior to processing of the multi-dimensional sub-volumes of the second input migrated seismic volume by the trained convolutional neural network.

16. The method of claim 10 , further comprising:

statistically aggregating a plurality of probability values produced by processing the multi-dimensional sub-volumes; and

applying a decision function to a result of the statistically aggregating to determine a state of a pixel corresponding to the plurality of probability values.

17. The method of claim 10 , further comprising:

applying a decision function to a plurality of probability values produced by processing the multi-dimensional sub-volumes; and

determining a state of a pixel corresponding to the plurality of probability values based on a result of the decision function for the plurality of probability values.

18. The method of claim 10 , further comprising:

simultaneously training a plurality of instances of the volumetric convolutional neural network using the first input migrated seismic volume and different hyperparameter values in each of the plurality of instances;

after completion of the simultaneous training, comparing accuracy of identification of the subsurface structure by each of the plurality of instances; and

selecting an instance of the plurality of instances that most accurately identifies the subsurface structure as the trained convolutional neural network.

19. The method of claim 10 , wherein a first multi-dimensional sub-volume overlaps a second multi-dimensional sub-volume by no more than 30%.

Assignments (1)

Continuity (1)

References Cited (34)