3D Seismic Velocity Estimation with Encoder-Decoder Networks
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing seismic velocity estimation methods, particularly those using deep learning, are limited to two-dimensional models and struggle with practical applications due to data complexity and computational costs, failing to accurately predict 3D subsurface velocities necessary for effective hydrocarbon reservoir identification and project planning.
Innovation Solution
A deep neural network trained on 5D seismic data, utilizing a 3D Encoder-Decoder architecture, learns to predict 3D velocity models directly from seismic shot records, incorporating data preprocessing and augmentation techniques to handle various survey types and geological complexities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional iterative algorithms (seismic tomography or Full Waveform Inversion) are used for velocity estimation, then velocity model accuracy is improved, but computational time and cost increase excessively
Solution Approach 1:
The patent replaces the mechanical iterative computational system with a neural network-based system. The neural network is trained on synthetic seismic data to learn the mapping from seismic data to velocity models, then applies this learned mapping to field data in a single forward pass, eliminating the need for iterative forward modeling and inversion calculations.
Solution Approach 2:
The patent performs preliminary training of the neural network model using extensive synthetic seismic data and corresponding velocity models before deployment. This preliminary action pre-computes the complex relationships between seismic data and velocity structures, allowing rapid inference on actual field data without repeating the computationally expensive iterative process.
2Loss of time
If deep learning algorithms are used for velocity prediction, then computational time is reduced, but the models are limited to 2D space and lack practical applicability
Solution Approach 1:
The patent extends deep learning from 2D to 3D velocity modeling by processing 3D seismic volumes and generating 3D velocity models. The neural network architecture handles three-dimensional spatial relationships and outputs three-dimensional velocity structures, enabling practical application to real 3D seismic survey data while maintaining the computational efficiency of direct prediction.
3Use of energy by moving object
If deep learning models are trained on 2D synthetic data, then training computational cost is reduced, but the models fail to accurately predict 3D subsurface velocities
Solution Approach 1:
The patent trains the neural network on 3D synthetic seismic data and corresponding 3D velocity models, matching the dimensionality of the target application. This 3D-to-3D training approach ensures the model learns accurate three-dimensional velocity structures and can generalize to field data, eliminating the dimensional mismatch that plagues 2D training approaches.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A method is described for estimating seismic velocity from seismic data by training a neural network using a subset of a seismic dataset and the velocity model; estimating a second velocity model using the neural network and a second subset of the seismic dataset; and displaying the second velocity model on a graphical user interface. The method may be executed by a computer system.