Deep Neural Network Initial Models for Seismic Least-Squares Migration
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional least-squares migration (LSM) methods are computationally expensive and inefficient, requiring numerous iterations and significant computational resources to achieve high-resolution seismic images, especially when dealing with high-frequency data and complex subsurface structures, which limits their practical application in hydrocarbon prospecting.
Innovation Solution
The implementation of deep neural networks, specifically reflectivity prediction networks and high-frequency prediction networks, to generate initial models that reduce the number of iterations needed for convergence, improving computational efficiency and image resolution by mapping input migrated images to estimates of true reflectivity and predicting high-frequency signals from low-frequency images.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional least-squares migration methods are used to achieve high-resolution seismic imaging, then image resolution is improved, but computational cost and processing time increase dramatically
Solution Approach 1:
The patent applies preliminary action by using deep neural networks to generate high-quality initial velocity models before the least-squares migration process begins. These pre-generated initial models significantly reduce the number of iterations required for convergence, thereby reducing computational time and processing costs while maintaining high image resolution.
Solution Approach 2:
The patent substitutes the traditional iterative mechanical optimization process with a deep neural network-based predictive system. The neural network learns from training data to directly predict improved initial models, replacing the conventional trial-and-error iterative approach with a more efficient intelligent system that reduces computational burden.
2Measurement precision
If conventional least-squares migration methods are used to achieve high-resolution seismic imaging, then image resolution is improved, but computational resources and cost increase significantly
Solution Approach 1:
By performing preliminary model generation using deep neural networks, the patent reduces the computational workload of the subsequent least-squares migration process. The pre-computed initial models require fewer iterations to converge, thereby reducing energy consumption and computational costs while achieving the same high-resolution imaging goal.
Solution Approach 2:
The patent replaces the computationally intensive iterative optimization mechanism with a deep neural network-based predictive mechanism. This substitution significantly reduces the computational resources and energy required to achieve high-resolution seismic images.
3Productivity
If standard depth migration algorithms are used, then processing speed is maintained, but image quality suffers from low resolution, uneven amplitude, and migration artifacts
Solution Approach 1:
The patent applies preliminary action by generating optimized initial velocity models using deep neural networks before performing depth migration. This preliminary step ensures that the subsequent migration process produces high-quality images with improved resolution, balanced amplitudes, and reduced artifacts, while maintaining reasonable processing speeds.
Solution Approach 2:
The patent substitutes the standard depth migration algorithm with an enhanced version that incorporates deep neural network predictions. This hybrid approach replaces the limitations of conventional algorithms with an intelligent system that simultaneously improves image quality and maintains processing efficiency.
4Measurement precision
If high frequencies are used in RTM to improve resolution, then image resolution is improved, but computational cost increases dramatically
Solution Approach 1:
The patent applies preliminary action by using deep neural networks to generate high-quality initial models that are optimized for high-frequency data processing. This preliminary optimization reduces the number of iterations required when processing high-frequency data, thereby reducing computational costs while maintaining the resolution benefits of high-frequency RTM.
Solution Approach 2:
The patent substitutes the computationally expensive high-frequency RTM process with a hybrid approach that uses deep neural network predictions to guide the migration. This substitution reduces the computational energy required to process high-frequency data while preserving the high-resolution imaging capability.
Data Source
AI summary
A method and apparatus for generating a high-resolution seismic image, including extracting a reflectivity distribution from a geological model; utilizing the reflectivity distribution to label features of the model; generating forward-modeled data from the model; migrating the forward-modeled data to create a migrated image; and training a deep neural network with the labeled synthetic geological model and the migrated image to create a reflectivity prediction network. A method and apparatus includes: selecting a first subset of the field data; applying a low-pass filter to the first subset to generate a first filtered dataset; migrating the first filtered dataset to create a first migrated image; applying a high-pass filter to the first subset to generate a second filtered dataset; migrating the second filtered dataset to create a second migrated image; and training a deep neural network to predict a target distribution of high-frequency signal.