Neural Network Seismic Velocity Inversion for Lower-Cost Imaging
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional methods for building accurate velocity models in seismic imaging require a large amount of seismic data and high computational costs, forcing a trade-off between model quality and cost.
Innovation Solution
A method using trainable travel time-based networks to construct velocity models, involving a trainable velocity network and a trainable travel time network, trained with a cost function that minimizes travel time mismatches and enforces travel time equations, to generate velocity models efficiently.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If conventional methods (ray-tracing tomography or full waveform inversion) are used to build an accurate velocity model, then the quality of the velocity model is improved, but the computational cost increases
Solution Approach 1:
The patent replaces conventional mechanical/computational seismic inversion methods (ray-tracing tomography and full waveform inversion) with a machine learning-based neural network system. The neural network is trained to directly predict velocity models from seismic data, substituting the complex iterative computational mechanics with a trained predictive model that achieves comparable accuracy at lower computational cost
Solution Approach 2:
The patent performs preliminary training of the neural network using a large dataset of seismic data and corresponding velocity models before actual velocity model building. This preliminary action creates a pre-trained model that can quickly generate velocity models without requiring the full computational power of conventional methods during the actual inversion process
2Measurement precision
If conventional methods are used with high computational cost, then an accurate velocity model is achieved, but the time required for processing increases
Solution Approach 1:
The patent substitutes time-consuming iterative inversion algorithms with a trained neural network that performs velocity model building in a single forward pass. The neural network architecture processes seismic data through multiple layers to directly output velocity models, eliminating the iterative loops that characterize conventional methods and dramatically reducing processing time
3Manufacturing precision
If conventional methods are used to build velocity models, then accurate seismic images can be produced, but the complexity of the system increases
Solution Approach 1:
The patent creates a universal neural network system that can handle multiple seismic inversion tasks through a single trained model. The network is designed to process different types of seismic data and generate velocity models for various geological conditions without requiring separate specialized algorithms for each case, thereby reducing overall system complexity
Solution Approach 2:
The patent extracts the essential velocity model building function from the complex conventional inversion process and encapsulates it in a trained neural network. By separating the core inversion logic into a self-contained trained model, the system eliminates the need for complex physical modeling, ray-tracing algorithms, and iterative optimization routines that characterize conventional methods
Data Source
AI summary
A method for training travel time-based networks and building an image of a velocity model includes obtaining a seismic dataset of seismic traces and determining an observed travel time for each seismic trace. The method further includes obtaining a velocity network, that depends on one or more velocity parameters, and a travel time network, that depends on one or more travel time parameters. The method further includes training the velocity network and the travel time network using a cost function and an optimizer. The cost function is based on the travel times parameters, the velocity parameters, a travel times equation, a derivative of the travel times equation, and a travel time mismatch between a first observed travel time and a first travel time value output by the travel time network. The method further includes building the image of a velocity model using the trained velocity network.