Granted Patent B2

US 12,417,382 · App. 17/195,755 · Granted Sep 16, 2025

Inventors: Rajesh S. Nair (Houston, TX); Karim Shafik Zaki (Houston, TX); Yan Li (Houston, TX); Margaretha Catharina Maria Rijken (Houston, TX); Velimir Valentinov Vesselinov (Los Alamos, NM)

Assignees: Chevron U.S.A. Inc.; TRIAD National Security, LLC

G06N3/08G06F30/23G06F30/27G06N3/04

Abstract

Claims (24)

1. A computer implemented method for prediction of productivity index decline for a hydrocarbon reservoir using a geomechanics informed machine learning (GIML) model, the method comprising:

performing geomechanical reservoir simulations to generate training datasets for the hydrocarbon reservoir;

incorporating identified variables into an initial layer of a neural network of the GIML model and intermediate variables into at least one intermediate layer of the neural network of the GIML model, wherein the intermediate variables are generated by applying total stress, strain-displacement, and elastic stress-strain calculations to the identified variables;

training the neural network of the GIML model with correlated training datasets that correlate to the physical models to produce a resulting prediction model, wherein the correlated training datasets are a subset of the training datasets, and wherein the training comprises training the initial layer of the neural network with the identified variables and training the at least one intermediate layer of the neural network with the intermediate variables;

performing sensitivity analysis on the resulting prediction model;

identifying dominant variables for damage mechanisms through design of experiment (DOE), wherein the damage mechanisms relate to the productivity index decline for the hydrocarbon reservoir;

performing history matching and blind test on the resulting prediction model;

predicting the productivity index decline for the hydrocarbon reservoir using the resulting prediction model; and

updating, based on an accuracy of the prediction of the productivity index decline being less than a threshold value, the identified variables and the intermediate variables incorporated into the GIML model.

2. The computer implemented method of claim 1 , wherein the GIML model is configured to calculate a productivity decline curve for each combination of the identified variables and the intermediate variables.

3. The computer implemented method of claim 1 , wherein the intermediate variables are further generated by applying to the identified variables geomechanics principles that comprise flow and transport principles, compaction and creep principles, fracture conductivity principles, perforation efficiency principles, and fines migration principles.

4. The computer implemented method of claim 1 , wherein the physical models are for prediction of the productivity index decline and comprise a fully coupled flow and geomechanics model used to simulate damage mechanisms during production of the hydrocarbon reservoir.

5. The computer implemented method of claim 4 , wherein the damage mechanisms comprise off-plane perforation contribution and stability, fracture connectivity and tortuosity, drilling and completion fluids invasion, creep and compaction effects, fracture conductivity, and fines migration and trapping.

6. The computer implemented method of claim 1 , wherein the identified variables are propagated through layers of the neural network of the GIML model to incorporate material models.

7. The computer implemented method of claim 1 , wherein the history matching is performed between field data for the hydrocarbon reservoir and the resulting prediction model, wherein the history matching generates results for identifying the variables and the physical models that have a larger impact on performance of the hydrocarbon reservoir then other variables.

8. A computer implemented method for predicting a geomechanical characteristic of a hydrocarbon reservoir using a geomechanics informed machine learning (GIML) model, the method comprising:

performing geomechanical reservoir simulations to generate training data;

incorporating identified variables into an initial layer of a neural network of the GIML model-and intermediate variables into at least one intermediate layer of the neural network of the GIML model, wherein the intermediate variables are generated by applying total stress, strain-displacement, and elastic stress-strain calculations to the identified variables;

training the neural network of the GIML model with correlated training data that correlates to the physical models to produce a resulting prediction model, wherein the correlated training data is a subset of the training data, and wherein the training comprises training the initial layer of the neural network with the identified variables and training the at least one intermediate layer of the neural network with the intermediate variables;

performing sensitivity analysis and history matching on the resulting prediction model;

predicting the geomechanical characteristic of the hydrocarbon reservoir using the resulting prediction model;

evaluating an accuracy of the geomechanical characteristic predicted using the resulting prediction model; and

updating the physical models based on the accuracy.

9. The computer implemented method of claim 8 , wherein training the neural network of the GIML model further comprises propagating the identified variables through layers of the neural network to incorporate material models.

Assignments (2)

Continuity (1)

References Cited (46)