Neural Network Interpolation for Seismic Data Densification

Solution Overview

Problem

Seismic data processing is hindered by suboptimal spatial sampling, leading to spatial aliasing and poor structural images, and existing methods for interpolating traces are inefficient and inaccurate, especially for real and simulated data.

Innovation Solution

Training neural networks using datasets with dense traces and subsets of coarse spatial sampling to interpolate and densify seismic data, enabling guided denoising and improving processing efficiency and accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If interpolation is applied to densify traces with coarse spatial sampling, then spatial sampling density is improved, but processing time and computational complexity increase

Engineering Contradiction:

Improvespatial sampling densityVSAvoidprocessing time

Core Design Contradiction:

Measurement precisionVSLoss of time

Solution Approach 1:

The neural network is trained in advance on a subset of data with known dense traces to learn the mapping from coarse to fine spatial sampling. This preliminary training enables the network to rapidly interpolate missing traces during actual processing without requiring intensive computational resources at that stage, thus resolving the contradiction between achieving high spatial sampling density and maintaining processing efficiency

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The neural network creates interpolated copies of existing traces to densify the spatial sampling. Instead of computing all traces from scratch or using computationally intensive interpolation methods, the network generates approximate copies of missing traces based on patterns learned from training data, significantly reducing processing time while maintaining acceptable spatial sampling density

Inventive Principle:

Principle #26Copying

2Object-affected harmful factors

If interpolation is applied to densify traces, then spatial aliasing is reduced, but manufacturing precision of the final structural image deteriorates due to inaccurate interpolation

Engineering Contradiction:

Improvespatial aliasingVSAvoidstructural image accuracy

Core Design Contradiction:

Object-affected harmful factorsVSManufacturing precision

Solution Approach 1:

The neural network is trained in advance on a subset of data with known dense traces to learn the mapping from coarse to fine spatial sampling. This preliminary training enables the network to rapidly interpolate missing traces during actual processing without requiring intensive computational resources at that stage, thus resolving the contradiction between achieving high spatial sampling density and maintaining processing efficiency

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The neural network uses feedback from the training process where it learns from the relationship between coarse and dense traces. The network adjusts its internal parameters based on the error between predicted and actual dense traces, enabling it to produce more accurate interpolations that reduce spatial aliasing while preserving structural image fidelity

Inventive Principle:

Principle #23Feedback

3Productivity

If simulated data is used for training, then training efficiency is improved, but reliability of the interpolation deteriorates due to simulation artifacts

Engineering Contradiction:

Improvetraining efficiencyVSAvoidinterpolation accuracy

Core Design Contradiction:

ProductivityVSReliability

Solution Approach 1:

The neural network is trained in advance on a subset of data with known dense traces to learn the mapping from coarse to fine spatial sampling. This preliminary training enables the network to rapidly interpolate missing traces during actual processing without requiring intensive computational resources at that stage, thus resolving the contradiction between achieving high spatial sampling density and maintaining processing efficiency

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The approach changes the parameter of data authenticity by using simulated data for training (which is computationally efficient) while applying the trained model to real data (which provides reliability). The simulated data allows rapid training with controlled parameters, and the learned interpolation patterns are then validated and applied to real seismic data, combining the benefits of both simulated and real data approaches

Inventive Principle:

Principle #35Parameter changes

Data Source

PatentUS11796700B2Using neural networks for interpolating seismic data

Publication Date: 2023.10.24 CGG SERVICES SAS

AI summary

One method interpolates simulated seismic data of a coarse spatial sampling to a finer spatial sampling using a neural network. The neural network is previously trained using a set of simulated seismic data with the finer spatial sampling and a subset thereof with the coarse spatial sampling. The data is simulated using an image of the explored underground formation generated using real seismic data. The seismic dataset resulting from simulation and interpolation is used for denoising the seismic data acquired over the underground formation. Another method demigrates seismic data at a sparse density and then increases density by interpolating traces using a neural network.