US-12625297-B2 - Automatic subsurface property model building and validation

Abstract

A method for modeling a subsurface property for a subterranean volume of interest includes receiving input measurement data representing a subterranean volume of interest, predicting a subsurface property based at least in part on the input measurement data using a first machine learning model, predicting a subsurface property model based at least in part on the subsurface property, the input measurement data, or both, using a second machine learning model, predicting synthetic measurement data based at least in part on the subsurface property model using a third machine learning model, a physics-based model, or both, comparing the synthetic measurement data and the input measurement data, and training the first machine learning model, the second machine learning model, or both based at least in part on the comparing.

Inventors

• Aria Abubakar
• Haibin Di
• Tao Zhao
• Zhun LI
• Cen Li

Assignees

• SCHLUMBERGER TECHNOLOGY CORPORATION

Dates

Publication Date

20260512

Application Date

20220322

Claims (18)

1. 1 . A method for modeling a subterranean volume of interest, comprising: receiving input measurement data representing the subterranean volume of interest; predicting a subsurface property based at least in part on the input measurement data using a first machine learning model, wherein the first machine learning model includes a self-supervised learning model and a supervised learning model, wherein the self-supervised learning model including a self-supervised learning (SSL) encoder and a SSL decoder, and the supervised learning model includes the SSL decoder and a supervised learning (SL) encoder, and wherein predicting the subsurface property includes: combining seismic traces with noise; training the self-supervised learning model to reconstruct the seismic traces from the seismic traces combined with the noise, resulting in reconstructed seismic traces; displaying the reconstructed seismic traces; training the supervised learning model to predict one or more properties based on the reconstructed seismic traces; and predicting the subsurface property using the supervised learning model; predicting a subsurface property model based at least in part on the subsurface property, the input measurement data, or both, using a second machine learning model; predicting synthetic measurement data based at least in part on the subsurface property model using a third machine learning model, a physics-based model, or both; generating a reconstruction error of the synthetic measurement data by comparing the synthetic measurement data and the input measurement data; using the reconstruction error and the synthetic measurement data, re-training the self-supervised learning model and the supervised learning model to reconstruct the seismic traces, resulting in retrained reconstructed seismic traces; and displaying the retrained reconstructed seismic traces.
2. 2 . The method of claim 1 , wherein training the first machine learning model, the second machine learning model, or both comprises generating a training gradient using a fourth machine learning model.
3. 3 . The method of claim 1 , wherein the input measurement data comprises a well log, seismic data, or both.
4. 4 . The method of claim 1 , further comprising processing the input measurement data prior to predicting the subsurface property based at least in part on the input measurement data.
5. 5 . The method of claim 1 , wherein the subsurface property is selected from the group consisting of resistivity, density, one or more seismic attributes, relative geological time, fault location, sale body location, and stratigraphy.
6. 6 . The method of claim 1 , wherein the subsurface property model comprises a three-dimensional model of the subsurface property of the subterranean volume of interest, and wherein the subsurface property is selected from the group consisting of density, sonic wave velocity, acoustic impedance, porosity, and saturation.
7. 7 . The method of claim 1 , wherein training the first machine learning model comprises holding the second machine learning model constant while adjusting the first machine learning model based on a comparison between the synthetic measurement data and the input measurement data.
8. 8 . A system, comprising: one or more processors; and a memory system storing instructions that, when executed by at least one of the one or more processors, cause the system to perform operations, the operations comprising: receiving input measurement data representing a subterranean volume of interest; predicting a subsurface property based at least in part on the input measurement data using a first machine learning model, wherein the first machine learning model includes a self-supervised learning model and a supervised learning model, wherein the self-supervised learning model including a self-supervised learning (SSL) encoder and a SSL decoder, and the supervised learning model includes the SSL decoder and a supervised learning (SL) encoder, and wherein predicting the subsurface property includes: combining seismic traces with noise; training the self-supervised learning model to reconstruct the seismic traces from the seismic traces combined with the noise, resulting in reconstructed seismic traces; displaying the reconstructed seismic traces; training the supervised learning model to predict one or more properties based on the reconstructed seismic traces; and predicting the subsurface property using the supervised learning model; predicting a subsurface property model based at least in part on the subsurface property, the input measurement data, or both, using a second machine learning model; predicting synthetic measurement data based at least in part on the subsurface property model using a third machine learning model, a physics-based model, or both; generating a reconstruction error of the synthetic measurement data by comparing the synthetic measurement data and the input measurement data; using the reconstruction error and the synthetic measurement data, re-training the self-supervised learning model and the supervised learning model to reconstruct the seismic traces, resulting in retrained reconstructed seismic traces; and displaying the retrained reconstructed seismic traces.
9. 9 . The system of claim 8 , wherein training the first machine learning model, the second machine learning model, or both comprises generating a training gradient using a fourth machine learning model.
10. 10 . The system of claim 8 , wherein the input measurement data comprises a well log, seismic data, or both.
11. 11 . The system of claim 8 , wherein the operations further comprise processing the input measurement data prior to predicting the subsurface property based at least in part on the input measurement data.
12. 12 . The system of claim 8 , wherein the subsurface property is selected from the group consisting of resistivity, density, one or more seismic attributes, relative geological time, fault location, sale body location, and stratigraphy.
13. 13 . The system of claim 8 , wherein the subsurface property model comprises a three-dimensional model of the subsurface property of the subterranean volume of interest, and wherein the subsurface property is selected from the group consisting of density, sonic wave velocity, acoustic impedance, porosity, and saturation.
14. 14 . The system of claim 8 , wherein training the second machine learning model comprises holding the first machine learning model constant while adjusting the second machine learning model based on a comparison between the synthetic measurement data and the input measurement data.
15. 15 . A non-transitory computer-readable medium storing instructions that, when executed by at least one processor of a computing system, cause the computing system to perform operations, the operations comprising: receiving input measurement data representing a subterranean volume of interest; predicting a subsurface property based at least in part on the input measurement data using a first machine learning model, wherein the first machine learning model includes a self-supervised learning model and a supervised learning model, wherein the self-supervised learning model including a self-supervised learning (SSL) encoder and a SSL decoder, and the supervised learning model includes the SSL decoder and a supervised learning (SL) encoder, and wherein predicting the subsurface property includes: combining seismic traces with noise; training the self-supervised learning model to reconstruct the seismic traces from the seismic traces combined with the noise, resulting in reconstructed seismic traces; displaying the reconstructed seismic traces; training the supervised learning model to predict one or more properties based on the reconstructed seismic traces; and predicting the subsurface property using the supervised learning model; predicting a subsurface property model based at least in part on the subsurface property, the input measurement data, or both, using a second machine learning model; predicting synthetic measurement data based at least in part on the subsurface property model using a third machine learning model, a physics-based model, or both; generating a reconstruction error of the synthetic measurement data by comparing the synthetic measurement data and the input measurement data; using the reconstruction error and the synthetic measurement data, re-training the self-supervised learning model and the supervised learning model to reconstruct the seismic traces, resulting in retrained reconstructed seismic traces; and displaying the retrained reconstructed seismic traces.
16. 16 . The medium of claim 15 , wherein training the first machine learning model, the second machine learning model, or both comprises generating a training gradient using a fourth machine learning model.
17. 17 . The medium of claim 15 , wherein the operations further comprise processing the input measurement data prior to predicting the subsurface property based at least in part on the input measurement data.
18. 18 . The medium of claim 15 , wherein the subsurface property model comprises a three-dimensional model of the subsurface property of the subterranean volume of interest, and wherein the subsurface property is selected from the group consisting of density, sonic wave velocity, acoustic impedance, porosity, and saturation.

Description

CROSS-REFERENCE TO RELATED APPLICATIONS This application is a National Stage Entry of International Application No. PCT/US2022/071266, filed Mar. 22, 2022, which claims priority to U.S. Provisional Patent Application No. 63/164,259 which was filed on Mar. 22, 2021 and is incorporated herein by reference in its entirety. BACKGROUND Measurements are collected in oilfield and other contexts that can be used to calculate or “invert” various properties of a subterranean volume of interest. Various processing and modeling techniques are employed to perform such inversion. Generally, the inversion process relies on a combination of several different such processing and/or modeling techniques, with human intervention called for along the way. For example, subject matter experts may receive the data generated at one stage and interpret the data, e.g., labeling stratigraphies or structures/volumes of interests, etc., before passing the data to the next stage. The inclusion of such human-based efforts, as well as the disjointed nature of the process, may make the process time intensive, prone to errors, and expensive. SUMMARY Embodiments of the disclosure include a method for modeling a subsurface property for a subterranean volume of interest. The method includes receiving input measurement data representing the subterranean volume of interest, predicting a subsurface property based at least in part on the input measurement data using a first machine learning model, predicting a subsurface property model based at least in part on the subsurface property, the input measurement data, or both, using a second machine learning model, predicting synthetic measurement data based at least in part on the subsurface property model using a third machine learning model, a physics-based model, or both, comparing the synthetic measurement data and the input measurement data, and training the first machine learning model, the second machine learning model, or both based at least in part on the comparing. In an embodiment, training the first machine learning model, the second machine learning model, or both includes generating a training gradient using a fourth machine learning model. In an embodiment, the input measurement data includes a well log, seismic data, or both. In an embodiment, the method further includes processing the input measurement data using a machine learning model, a physics-based model, or both prior to predicting the subsurface property based at least in part on the input measurement data. In an embodiment, the subsurface property is selected from the group consisting of: resistivity, density, one or more seismic attributes, relative geological time, fault location, sale body location, and stratigraphy. In an embodiment, the subsurface property model includes a three-dimensional model of a property of the subterranean volume of interest, and the property is selected from the group consisting of density, sonic wave velocity, acoustic impedance, porosity, and saturation. In an embodiment, training the first machine learning model, the second machine learning model, or both includes holding the first machine learning model constant while adjusting the second machine learning model based on a comparison between the synthetic measurement data and the input measurement data, and holding the second machine learning model constant while adjusting the first machine learning model based on a comparison between the synthetic measurement data and the input measurement data. In an embodiment, predicting the subsurface property based at least in part on the input measurement data using a first machine learning model, includes combining seismic traces with noise, training a self-supervised learning model, including a SSL encoder and a SSL decoder, to reconstruct the seismic traces from the combination of the seismic traces and noise, training a supervised learning model, including the SSL decoder of the trained self-supervised learning model, and a SL encoder, to predict one or more properties based on seismic traces, and predicting the subsurface property using the supervised learning model. Embodiments of the disclosure further include a computing system including one or more processors and a memory system including one or more non-transitory computer-readable media storing instructions that, when executed by at least one of the one or more processors, cause the computing system to perform operations. The operations include receiving input measurement data representing a subterranean volume of interest, predicting a subsurface property based at least in part on the input measurement data using a first machine learning model, predicting a subsurface property model based at least in part on the subsurface property, the input measurement data, or both, using a second machine learning model, predicting synthetic measurement data based at least in part on the subsurface property model using a third machine learning model, a phy