Seismic Bandwidth Extension via Neural Network Synthetic Training

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Seismic datasets in oil and gas exploration have limited bandwidth, leading to inaccurate depiction of subsurface features and noise sensitivity, which hinders the interpretation of subsurface reflectors and hydrocarbon reservoir location.

Innovation Solution

A method using machine-learned models, specifically convolutional neural networks, to generate synthetic datasets with extended bandwidth by combining geometric shapes and wavelets, allowing for the training of models that can enhance seismic datasets to include lower frequencies, thereby improving the accuracy of subsurface modeling.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional seismic data acquisition is used, then the seismic dataset can be obtained, but the bandwidth is limited and cannot include sufficient lower frequency content

Engineering Contradiction:

ImprovebandwidthVSAvoidacquisition efficiency

Core Design Contradiction:

Measurement precisionVSProductivity

Solution Approach 1:

The patent generates synthetic bandwidth-extended seismic datasets in advance to serve as training data for machine learning models. These pre-generated datasets with extended low-frequency content enable the model to learn frequency extension patterns before actual seismic data processing, resolving the contradiction by preparing enhancement capabilities beforehand rather than requiring time-consuming field acquisition for extended bandwidth

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent replaces physical seismic data acquisition mechanisms with machine learning-based frequency extension. Instead of using conventional seismic sources and receivers to acquire broad bandwidth data, the system uses trained neural networks to synthesize extended bandwidth seismic datasets from limited-frequency input data, substituting mechanical acquisition with computational generation

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

2Measurement precision

If the bandwidth is extended to include lower frequencies, then the accuracy of subsurface modeling is improved, but the complexity of data processing increases

Engineering Contradiction:

Improvesubsurface modeling accuracyVSAvoiddata processing complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent introduces machine learning models as intermediaries between input seismic data and final subsurface models. The trained models act as mediators that automatically perform frequency extension and noise filtering, simplifying the processing pipeline while achieving enhanced bandwidth and improved subsurface imaging accuracy without requiring complex manual processing steps

Inventive Principle:

Principle #24Intermediary (Mediator)

Solution Approach 2:

The patent changes the frequency domain parameters of seismic data by extending bandwidth to include lower frequencies. The machine learning models transform the spectral characteristics of input data, modifying frequency content parameters to achieve broader bandwidth while maintaining processing efficiency through learned transformations rather than complex computational methods

Inventive Principle:

Principle #35Parameter changes

3Ease of manufacture

If the seismic dataset has limited bandwidth, then the data acquisition is simpler, but the depiction of subsurface features is inaccurate and noise-sensitive

Engineering Contradiction:

Improvedata acquisition simplicityVSAvoidsubsurface feature depiction accuracy

Core Design Contradiction:

Ease of manufactureVSReliability

Solution Approach 1:

The patent creates synthetic copies of seismic data with extended bandwidth characteristics. By generating synthetic datasets that mimic the statistical and spectral properties of real seismic data but with enhanced low-frequency content, the system preserves the simplicity of acquiring limited-bandwidth field data while producing reliable subsurface images through computational copying and enhancement

Inventive Principle:

Principle #26Copying

Data Source

PatentUS20240329266A1Robust low frequency seismic bandwidth extension via a deep neural network trained on synthetic seismic data

Publication Date: 2024.10.03 SAUDI ARABIAN OIL CO

AI summary

Systems and methods are disclosed. The method includes collecting a seismic dataset from a seismic survey and generating a plurality of synthetic datasets each including an input seismic dataset with a first bandwidth and an associated target seismic dataset with a second, broader, bandwidth, that includes lower frequencies than the first bandwidth. Each synthetic dataset is generated in a time domain and includes an event with a geometric shape and a wavelet. The method also includes splitting the plurality of synthetic datasets into a training set; selecting and training a first machine-learned model with a first architecture, to receive the seismic dataset and output another seismic dataset with extended bandwidth relative to the seismic dataset. The method further includes using the first machine-learned model to produce an extended bandwidth seismic dataset from the seismic dataset; and determining a location of a hydrocarbon reservoir using the extended bandwidth seismic dataset.