Density Interface Inversion via Machine Learning Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for inverting abrupt density interfaces in geophysical data, such as regularization inversion and gravity migration, face challenges in resolution and interpretability, with regularization inversion being vague and gravity migration having low resolution, while deep learning methods struggle with interpretability due to limited data sets and single input training modes.
Innovation Solution
A density abrupt interface inversion method based on machine learning constraints is developed, involving constructing an initial basin interface, generating disturbed data sets, performing Hadamard product operations, and optimizing a migration model deep learning network with depth weighting and multivariate density contraction constraints to enhance imaging resolution and interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If regularization inversion imaging is used to invert density interfaces, then the stability of inversion is improved through dynamic weight allocation, but the imaging resolution of abrupt density interfaces remains vague
Solution Approach 1:
The patent combines regularization inversion with deep learning migration imaging to merge the stability advantages of regularization inversion with the high-resolution capabilities of deep learning-based methods. The deep learning model is trained using both measured gravity data and simulated data from regularization inversion results, allowing the system to achieve both stability and high resolution in density interface imaging.
Solution Approach 2:
The patent introduces an intermediary deep learning migration imaging step between the gravity data and the final density interface model. This intermediary layer processes the gravity data through a trained deep learning model that has learned the relationship between gravity anomalies and density interfaces, producing high-resolution density models that bridge the gap between stable inversion and high-resolution imaging.
2Reliability
If gravity migration imaging is used to model density interfaces, then the stability and target shape recovery ability are improved, but the resolution is too low to meet geological interpretation needs
Solution Approach 1:
The patent replaces traditional mechanical migration imaging methods with a deep learning-based imaging approach. The deep learning model is trained to directly map gravity data to high-resolution density interfaces, substituting the conventional step-by-step mechanical migration process with a data-driven neural network that achieves both stability and high resolution simultaneously.
Solution Approach 2:
The patent changes the fundamental parameters of the imaging process by using deep learning models with multiple input channels (measured gravity data, simulated gravity data, and initial models) and adjusting the network architecture parameters (convolutional layers, pooling layers, activation functions) to optimize both resolution and stability for density interface imaging.
3Adaptability or versatility
If traditional deep learning imaging is used to predict density interfaces, then the generalization and feature extraction ability are improved, but the interpretability of the optimization process is affected due to limited data sets
Solution Approach 1:
The patent implements feedback mechanisms in the deep learning model by using simulated gravity data generated from regularization inversion results as input to the migration imaging model. This feedback loop allows the model to continuously refine its predictions by comparing simulated and measured gravity data, improving both the interpretability of the optimization process and the accuracy of density interface reconstruction.
Solution Approach 2:
The patent performs preliminary actions by pre-training the deep learning model using simulated data generated from regularization inversion results before applying it to real measured gravity data. This preliminary training phase establishes the model's interpretability and optimization process using controlled synthetic data, which then guides the interpretation of results from actual geological data.
Data Source
AI summary
Disclosed are a hybrid density abrupt interface inversion method based on machine learning constraints. The inversion method includes constructing an initial basin interface and randomly generating a disturbed basin interface data set; obtaining a basin interface data set through Hadamard product operation on the initial basin interface and the disturbed basin interface data set; obtaining a high-resolution density interface model data set through filling the basin interface data set with advanced functions; performing forward calculation to obtain a simulated gravity data set; carrying out mathematical transformation on the simulated gravity data set and weighting to obtain a low-resolution migration density interface model data set; optimizing a migration model-based deep learning network and mapping to obtain a high-resolution constrained density interface prior model; and constructing a stable nonlinear loss function and performing regularization inversion to obtain a high-resolution density interface model.