Neural Network Seismic Inversion Without Wavelet Estimation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional seismic inversion methods rely on wavelet estimation, which is complex, subjective, and time-consuming, introducing inaccuracies and biases, especially when dealing with spatially varying wavelets across multiple wells, thus limiting the efficiency and reliability of rock property estimation.
Innovation Solution
A new training strategy for neural networks that eliminates the need for wavelet estimation by using data augmentation to generate synthetic training datasets, allowing the network to learn from seismic and well log data without relying on wavelet knowledge, utilizing deep neural networks and convolutional neural networks to directly predict impedance profiles from seismic data.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional seismic inversion methods using wavelet estimation are used, then rock property estimation can be performed, but the process becomes complex, subjective, and time-consuming with introduced inaccuracies and biases
Solution Approach 1:
The patent extracts and removes the wavelet estimation step from the seismic inversion process. By training neural networks to directly map seismic traces to rock properties without requiring wavelet estimation, the method eliminates the source of subjectivity and complexity while maintaining or improving accuracy through data-driven learning
Solution Approach 2:
The patent replaces the traditional mechanical/mathematical wavelet estimation process with a data-driven neural network approach. The neural network learns the relationship between seismic data and rock properties directly from training data, substituting the complex analytical process with a trained model that provides consistent, objective results
2Measurement precision
If wavelet estimation is performed for each well, then accurate rock properties can be estimated, but the processing time increases significantly when dealing with multiple wells
Solution Approach 1:
The patent performs preliminary training of neural networks using synthetic data generated from a limited set of wells. Once trained, the network can rapidly process additional wells without requiring time-consuming wavelet estimation for each one, thus achieving both precision and efficiency
Solution Approach 2:
The patent uses synthetic training data that copies the essential characteristics of real seismic and well log data. This synthetic data allows the network to learn from diverse scenarios without requiring actual measurements from every possible well configuration, reducing the time needed for processing real wells
3Ease of manufacture
If traditional neural network training without data augmentation is used, then training can be completed with available data, but the model may overfit and fail to generalize to new seismic data
Solution Approach 1:
The patent performs preliminary data augmentation during the training phase, generating synthetic variations of the training data before the actual training process. This prepares a more robust training set that helps the model generalize better to unseen data while maintaining ease of implementation through automated synthetic data generation
Solution Approach 2:
The patent creates a composite training dataset that combines real seismic and well log data with synthetically augmented variations. This composite dataset provides both the authenticity of real measurements and the diversity of synthetic variations, improving model generalization without complicating the training process
Data Source
Figure 1
Figure 2
Figure 3A~3B
AI summary
A method is described for property estimation including receiving a seismic dataset representative of a subsurface volume of interest and a well log from a well location within the subsurface volume of interest; identifying seismic traces in the seismic dataset that correspond to the well location to obtain a subset of seismic traces; windowing the subset of seismic traces and the well log to generate windowed seismic traces and a windowed well log; multiplying the windowed seismic traces and the windowed well log by a random matrix to generate a plurality of training datasets; and training a neural network using the plurality of training datasets. The method may be executed by a computer system.