Machine learning inversion using bayesian inference and sampling

1 . A method comprising:

iteratively:

obtaining an observed geophysical dataset from a subterranean region of interest;

determining, using a computer processor, a preprocessed observed geophysical dataset based on the observed geophysical dataset;

obtaining a current geophysical model of the subterranean region of interest;

forming a training dataset comprising a plurality of geophysical training models and a corresponding simulated geophysical training dataset for each of the plurality of geophysical training models;

performing inversion comprising iteratively or recursively, using the computer processor, until a convergence criterion is satisfied:

determining a simulated geophysical dataset by applying forward modeling to the current geophysical model,

determining a data loss function between the preprocessed observed geophysical dataset and the simulated geophysical dataset,

training a machine learning (ML) network, using the training dataset, to predict a predicted geophysical model from the preprocessed observed geophysical dataset,

inputting the preprocessed observed geophysical dataset into a trained machine learning network,

wherein the trained machine learning network comprises Bayesian inference;

producing a predicted geophysical model from the trained machine learning network using the preprocessed observed geophysical dataset,

wherein the predicted geophysical model comprises a plurality of model parameters, and

wherein each model parameter among the plurality of model parameters comprises a probability distribution of each model parameter,

determining a model loss function between the current geophysical model and the predicted geophysical model,

forming an objective function based on the data loss function and the model loss function,

determining a value of the objective function, and

updating the current geophysical model based on the value; and

locating a hydrocarbon reservoir based on the updated current geophysical model;

planning a wellbore path to the hydrocarbon reservoir;

drilling, using a geosteering system, a wellbore guided by the wellbore path comprising updating a position of a drill bit of the geosteering system based on the updated current geophysical model;

determining, using the geosteering system and during drilling, acoustic measurements; and

updating the observed geophysical dataset using the acoustic measurements.

2 . The method of claim 1 , wherein determining the preprocessed observed geophysical dataset comprises sampling the observed geophysical dataset.

3 . The method of claim 1 , wherein training the machine learning network comprises sampling a training dataset.

4 . The method of claim 3 , wherein sampling comprises active learning.

5 . The method of claim 1 , wherein the Bayesian inference comprises a Markov chain Monte Carlo method.

6 . The method of claim 1 , wherein the machine learning network comprises Gaussian process regression.

7 . The method of claim 1 , wherein the observed geophysical dataset comprises a seismic dataset.

8 . The method of claim 1 , wherein the trained machine learning network comprises a supervised machine learning network.

9 . The method of claim 1 , wherein training the trained machine learning network comprises:

augmenting a training dataset with the updated current geophysical model and the simulated geophysical dataset; and

retraining the trained machine learning network based on the augmented training dataset.

10 . A system comprising:

a seismic acquisition system configured to obtain an observed geophysical dataset from a subterranean region of interest;

a computer system configured to, iteratively:

receive the observed geophysical dataset,

determine a preprocessed observed geophysical dataset based on the observed geophysical dataset,

receive a current geophysical model of the subterranean region of interest,

form a training dataset comprising a plurality of geophysical training models and a corresponding simulated geophysical training dataset for each of the plurality of geophysical training models;

perform inversion comprising iteratively or recursively, until a convergence criterion is satisfied:

determine a simulated geophysical dataset by applying forward modeling to the current geophysical model;

determine a data loss function between the preprocessed observed geophysical dataset and the simulated geophysical dataset;

train a machine learning (ML) network, using the training dataset, to predict a predicted geophysical model from the preprocessed observed geophysical dataset,

input the preprocessed observed geophysical dataset into a trained machine learning network,

wherein the trained machine learning network comprises Bayesian inference;

produce a predicted geophysical model from the trained machine learning network using the preprocessed observed geophysical dataset,

wherein the predicted geophysical model comprises a plurality of model parameters, and

wherein each model parameter among the plurality of model parameters comprises a probability distribution of each model parameter;

determine a model loss function between the current geophysical model and the predicted geophysical model;

form an objective function based on the data loss function and the model loss function;

determine a value of the objective function;

update the current geophysical model based on the value, and

locate a hydrocarbon reservoir based on the updated current geophysical model;

a wellbore planning system configured to plan a wellbore path to the hydrocarbon reservoir; and

a geosteering system configured to:

drill a wellbore guided by the wellbore path comprising update a position of a drill bit of the geosteering system based on the updated current geophysical model, and

determine, during drilling, acoustic measurements;

wherein the computer system is further configured to update the observed geophysical dataset using the acoustic measurements.