Geophysical Inversion Using Autoencoder Latent Representations

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current seismic prospecting methods, particularly Full Wavefield Inversion (FWI), face challenges in computational efficiency and accuracy due to non-uniqueness, lack of convexity in objective functions, and sensitivity to initial models, leading to inaccurate and inefficient subsurface modeling.

Innovation Solution

The implementation of machine learning-augmented geophysical inversion using autoencoders and generative-adversarial networks to learn data-space and model-space representations, which reformulates the objective function in a lower-dimensional space, mitigating non-uniqueness and improving the comparison of simulated and measured data.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If Full Wavefield Inversion (FWI) is used to model subsurface geologic structures, then detailed subsurface imaging is achieved, but computational cost and time consumption increase significantly

Engineering Contradiction:
Improvesubsurface imaging accuracyVSAvoidcomputational time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by using autoencoders to pre-process and compress seismic data into latent representations before inversion. This preprocessing step transforms the raw high-dimensional seismic data into a compressed form that captures essential features, reducing the computational burden of subsequent inversion operations while preserving the information needed for accurate subsurface imaging.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces an intermediary mechanism by incorporating learned representations from autoencoders as intermediate variables in the inversion process. These learned representations act as mediators between the raw seismic data and the subsurface model parameters, enabling the inversion to work with compressed, feature-rich data that requires less computational resources while maintaining imaging accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If traditional inversion procedures are used without incorporating geological knowledge, then computational simplicity is maintained, but solution accuracy and geological合理性 decrease due to non-uniqueness

Engineering Contradiction:
Improvesolution accuracyVSAvoidinversion procedure complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent applies feedback by integrating geological knowledge and constraints into the inversion objective function. The learned representations from autoencoders provide feedback about the expected geological structures and patterns, guiding the inversion toward solutions that are both data-consistent and geologically reasonable. This feedback mechanism helps resolve non-uniqueness by incorporating prior geological information into the optimization process.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The patent changes parameters by transforming the inversion problem into a different parameter space using the autoencoder's latent representations. Instead of directly inverting for subsurface parameters from raw seismic data, the method inverts for compressed representations that encode geological information, then reconstructs the final model. This parameter transformation reduces non-uniqueness while maintaining computational tractability.

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If uniform discretization is used for parameterization, then implementation simplicity is maintained, but model accuracy decreases due to inability to capture fine-scale variations

Engineering Contradiction:
Improvemodel accuracyVSAvoidparameterization complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent applies local quality by allowing the discretization and parameterization to vary locally based on geological features. The autoencoder learns to represent different regions of the subsurface with appropriate levels of detail, capturing fine-scale variations in complex areas while using coarser representation in simpler regions. This localized adaptation improves model accuracy without requiring uniformly fine discretization everywhere.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent changes dimensions by moving from direct physical space parameterization to a latent representation space. The autoencoder transforms the high-dimensional subsurface model into a lower-dimensional latent space that captures essential geological variations. This dimensional transformation enables the model to represent fine-scale features efficiently without requiring correspondingly fine physical discretization.

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

Data Source

PatentEP3894907B1Machine learning-augmented geophysical inversion
Publication Date: 2024.01.24 EXXONMOBIL TECHNOLOGY & ENGINEERING CO
  • EP3894907B1 patent drawingFigure 1
  • EP3894907B1 patent drawingFigure 2A
  • EP3894907B1 patent drawingFigure 2B

AI summary

A method and system of machine learning-augmented geophysical inversion includes obtaining measured data; obtaining prior subsurface data; (a) partially training a data autoencoder with the measured data to learn a fraction of data space representations and generate a data space encoder; (b) partially training a model autoencoder with the prior subsurface data to learn a fraction of model space representations and generate a model space decoder; (c) forming an augmented forward model with the model space decoder, the data space encoder, and a physics-based forward model; (d) solving an inversion problem with the augmented forward model to generate an inversion solution; and iteratively repeating (a) - (d) until convergence of the inversion solution, wherein, for each iteration: partially training the data and model autoencoders starts with learned weights from an immediately-previous iteration; and solving the inversion problem starts with super parameters from the previous iteration.