Backpropagation Neural Network for Subsurface Fluid Identification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional geophysical inversion methods face challenges such as computational inefficiency and non-uniqueness of solutions when identifying subsurface fluids and lithologies, as they rely on traditional approaches and require large amounts of data to train, limiting their scalability and accuracy.
Innovation Solution
A method utilizing a backpropagation-enabled process, specifically deep learning techniques like convolutional neural networks, is employed to train on a geophysical data set for inferring the presence of subsurface pore-filling fluids and lithologies, incorporating supervised, semi-supervised, or unsupervised learning approaches to improve accuracy and efficiency while reducing computational resource needs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional geophysical inversion methods are used, then the process can identify subsurface fluids and lithologies, but the computational efficiency is poor and the solutions are non-unique
Solution Approach 1:
The patent replaces the mechanical iterative inversion process with a neural network-based system. The neural network is trained on synthetic seismic data to learn the mapping from seismic responses to subsurface properties, eliminating the need for traditional forward modeling and iterative inversion. This substitution provides unique solutions efficiently without the computational burden and non-uniqueness issues of conventional methods.
Solution Approach 2:
The patent performs preliminary training of the neural network using extensive synthetic data before actual subsurface identification. This pre-training phase creates a robust model that can directly predict subsurface properties from seismic data without requiring iterative refinement during the actual application, thereby achieving both uniqueness and computational efficiency.
2Reliability
If artificial neural networks are applied to geophysical inversion, then data-processing challenges are addressed, but large amounts of data are required for training and scalability is limited
Solution Approach 1:
The patent changes the parameters of the neural network architecture and training approach to reduce data requirements. By optimizing network structure and training methodology, the system achieves effective subsurface identification with reduced training data compared to conventional ANN applications, while maintaining robust data-processing capabilities.
3Device complexity
If deep convolutional networks are used, then the number of parameters to be learned is reduced, but the approach still relies on forward modeling and objective functions that may become trapped
Solution Approach 1:
The patent inverts the traditional approach by training the neural network to directly predict subsurface properties from seismic data rather than using forward modeling to predict seismic responses from subsurface models. This inversion eliminates the objective function that can become trapped in local minima, as the network learns the inverse mapping directly during training, ensuring more reliable convergence.
Data Source
AI summary
A method for a method for identifying a subsurface pore-filling fluid and/or lithology. A training set of field-acquired geophysical data and/or simulated geophysical data is provided to train a backpropagation-enabled process. The trained process is used on a field-acquired data set that is not part of the training set to infer presence of a subsurface pore-filling fluid and/or lithology.