GAN-Based UDD-to-MDD Seismic Imaging for Stable Green's Function

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing data-dependent re-datuming techniques for estimating Green's function are unstable due to the presence of seabed tilt with sharp lateral variations, making up-down deconvolution (UDD) ineffective in complex subsurface structures.

Innovation Solution

A method involving a generative adversarial network (GAN) is used to transform upward-downward diffusion (UDD) seismic images into multi-dimensional deconvolution (MDD) images through iterative training of a generator and discriminator, enabling accurate mapping between the two domains.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If up-down deconvolution (UDD) is used to estimate Green's function, then the process is simple and fast, but the estimation becomes unstable in the presence of seabed tilt with sharp lateral variations

Engineering Contradiction:

Improveprocessing speedVSAvoidstability of Green's function estimation

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The patent introduces a deep learning model as an intermediary that transforms UDD seismic images into MDD seismic images. This mediator enables the system to leverage the simplicity of UDD while achieving the stability and accuracy of MDD by learning the complex transformation between the two deconvolution methods through trained neural network weights.

Inventive Principle:

Principle #24Intermediary (Mediator)

2Reliability

If multi-dimensional deconvolution (MDD) is used to estimate Green's function, then the estimation is stable and accurate, but the computational complexity increases

Engineering Contradiction:

Improvestability of Green's function estimationVSAvoidcomputational complexity

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent creates a learned copy of the MDD transformation process through a deep learning model. Instead of performing computationally intensive MDD calculations, the system trains a neural network to copy the essential transformation from UDD to MDD images, enabling fast inference that replicates MDD results without the original computational burden.

Inventive Principle:

Principle #26Copying

Solution Approach 2:

The patent performs preliminary action by pre-training the deep learning model on paired UDD and MDD seismic images before actual processing. This offline training phase computes the complex transformation relationships in advance, storing them in learned weights that can be rapidly applied during production without repeating the intensive MDD calculations.

Inventive Principle:

Principle #10Preliminary action

3Ease of manufacture

If UDD is applied in complex subsurface structures with seabed tilt, then processing remains simple, but the quality of seismic imaging deteriorates

Engineering Contradiction:

Improvesimplicity of processingVSAvoidquality of seismic imaging

Core Design Contradiction:

Ease of manufactureVSManufacturing precision

Solution Approach 1:

The patent changes the parameter representation by transforming UDD seismic images into MDD seismic images through deep learning. This parameter transformation allows the system to maintain the simple UDD processing workflow while achieving the superior imaging quality of MDD by learning the optimal parameter mapping between the two methods.

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentEP4686961A1Extrapolating green's function estimated using multidimensional deconvolution beyond receiver grid through deep learning

Publication Date: 2026.02.04 SERVICES PETROLIERS SCHLUMBERGER SA

Figure 1
Figure 2
Figure 3~4

AI summary

A method for transforming seismic images includes receiving input data. The input data includes an original upward-downward diffusion (UDD) seismic image and an original multi-dimensional deconvolution (MDD) seismic image. The method also includes training a generator and a discriminator based upon the input data to produce a trained generator and a trained discriminator.