3D ATEM Inversion Using U-Net and Random Resistivity Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current three-dimensional inversion methods for airborne transient electromagnetics face challenges such as high computational requirements, sensitivity to parameters, susceptibility to local minima, and limited applicability of deep learning due to lack of training samples, leading to poor inversion results and difficulty in real-time processing.
Innovation Solution
A three-dimensional inversion method using deep learning that decomposes a large-scale region into local models, trains and predicts them, and reconstructs a final model, employing a U-Net architecture with a mean square error loss function and Adam optimization, and generates training data through a random resistivity model algorithm.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If three-dimensional inversion is used to provide detailed underground resistivity distribution information, then measurement precision is improved, but computing time increases significantly to at least 10 hours for large datasets
Solution Approach 1:
The patent pre-trains a deep learning model using synthetic training data generated from forward modeling of three-dimensional electromagnetic responses. This preliminary training enables the model to learn complex nonlinear mappings between observed data and underground resistivity structures, allowing rapid inversion during actual application without performing computationally intensive iterative optimization in real-time.
Solution Approach 2:
The patent creates synthetic training data by generating copies of realistic underground resistivity models through forward modeling. These synthetic datasets replicate the statistical characteristics and physical relationships of real geological structures, enabling the deep learning model to learn from numerous examples without requiring actual field data for each inversion case.
2Manufacturing precision
If gradient-based methods are used for three-dimensional inversion optimization, then manufacturing precision is improved, but reliability deteriorates due to susceptibility to local minima
Solution Approach 1:
The patent replaces traditional gradient-based mechanical optimization methods with a deep learning-based intelligent system. The neural network learns the inverse mapping from data to model through training, eliminating the need for iterative gradient descent and its associated problems with local minima. The model directly predicts resistivity distributions based on learned patterns from training data.
3Measurement precision
If three-dimensional inversion with over 200,000 grids is performed to meet exploration needs, then measurement precision is improved, but device complexity increases enormously
Solution Approach 1:
The patent performs the computationally intensive task of exploring the solution space during the offline training phase, where the deep learning model learns optimal inversion strategies for various geological scenarios. During actual inversion operations, the pre-trained model provides rapid predictions without requiring complex iterative computations, effectively transferring the complexity burden from runtime to training time.
4Productivity
If deep learning is applied to three-dimensional inversion without sufficient training samples, then productivity is improved through automation, but measurement precision deteriorates due to lower data fitting performance
Solution Approach 1:
The patent performs the computationally intensive task of exploring the solution space during the offline training phase, where the deep learning model learns optimal inversion strategies for various geological scenarios. During actual inversion operations, the pre-trained model provides rapid predictions without requiring complex iterative computations, effectively transferring the complexity burden from runtime to training time.
Solution Approach 2:
The patent creates synthetic training data by generating copies of realistic underground resistivity models through forward modeling. These synthetic datasets replicate the statistical characteristics and physical relationships of real geological structures, enabling the deep learning model to learn from numerous examples without requiring actual field data for each inversion case.
Data Source
AI summary
The present disclosure discloses a three-dimensional inversion method of airborne transient electromagnetics based on deep learning. The method of the present disclosure proposes two strategies that focus on training datasets, to improve the performance of deep learning models, including divide and conquer strategy and random models generating. Through a large of reasonable structural models, appropriate network setups, a more generalized result can be obtained through our proposed U-Net framework, which has been demonstrated to be effective on both synthetic and field data. This scheme can realize the rapid prospecting of three-dimensional resistivity structure in large-area target region, and solve the problem of low efficiency of traditional three-dimensional inversion calculation of ATEM and poor migration ability of three-dimensional inversion based on deep learning developed by predecessors.