Low Frequency Model Generation for Seismic Inversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Seismic inversion in regions with limited control wells faces challenges in generating accurate low frequency models due to insufficient data, leading to uncertainty in determining rock properties and hydrocarbon exploration efforts.
Innovation Solution
A machine learning model is used to generate low frequency models based on seismic data and well log data from regions with control wells, which are then extrapolated to areas without control wells, recursively scaling the models to span larger regions, thereby improving accuracy and reducing bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional co-kriging methods are used to extrapolate low frequency models between wells, then the process can be performed with available data, but the accuracy of the low frequency model decreases in regions with limited control wells
Solution Approach 1:
The patent uses machine learning models as an intermediary between seismic data and well log data to generate low frequency models. The ML model learns relationships from training data in regions with control wells and applies this knowledge to regions without control wells, acting as a mediator that transfers information across data-scarce boundaries.
Solution Approach 2:
The patent performs preliminary training of machine learning models using seismic and well log data from regions with control wells before applying the models to regions without control wells. This preliminary action creates a knowledge base that can be recursively applied to expand the low frequency model across the entire domain.
2Reliability
If seismic inversion is performed without sufficient low frequency data, then the process can proceed with available seismic data, but the reliability of inverted rock properties decreases
Solution Approach 1:
The patent replaces traditional mechanical interpolation methods (co-kriging) with machine learning-based synthesis. The ML model synthesizes low frequency models by learning complex nonlinear relationships from training data, providing more reliable results than linear interpolation methods when data is sparse.
Solution Approach 2:
The patent changes the frequency parameter by generating low frequency models (typically 5-30 Hz) that complement the higher frequency seismic data. This parameter change addresses the information loss at low frequencies and improves the reliability of inverted rock properties.
3Area of stationary object
If low frequency models are generated for large regions without control wells, then the coverage area increases, but the accuracy and confidence in the model decreases
Solution Approach 1:
The patent segments the large region into smaller sub-regions and performs recursive extrapolation step-by-step. The ML model generates low frequency models for regions with control wells, then recursively extrapolates to adjacent regions without control wells, maintaining better accuracy through incremental expansion rather than single-step large-scale extrapolation.
Solution Approach 2:
The patent uses the spatial dimension recursively, expanding the low frequency model coverage progressively from regions with control wells to adjacent regions without control wells. This dimensional expansion approach maintains accuracy by building upon established model regions rather than extrapolating across entire domains at once.
Data Source
AI summary
The systems and methods described in this specification relate to generating a low frequency model of a subterranean formation for performing a seismic inversion. The systems and methods receive seismic data for a first region of the subterranean formation and well log data of one or more wells located at the first region. The systems and methods determine one or more relative layer attributes of the first region, one or more first input values for a machine learning model, and one or more second input values for the machine learning model. The systems and methods generate, a first relative low frequency model for the first region, and extrapolate, by executing the machine learning model by the processor, the first relative low frequency model to a second region of the subterranean formation.