Disentangled Factor Learning for Seismic Data Inference
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Reservoir modeling and simulation in the oil & gas industry face challenges due to incomplete data from seismic surveys and well logging, particularly in complex subsurface formations, requiring manual interpretation and lacking efficient automated methods for inferring subsurface properties.
Innovation Solution
A disentangled factor learning framework is applied to petro-technical image data, such as seismic data, to create disentangled representations of subsurface attributes, using algorithms like general adversarial networks and variational autoencoders, enabling the generation of training data for machine learning algorithms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If manual interpretation of seismic survey data is used to identify geological layers, then accuracy of subsurface structure determination is improved, but time consumption and operational complexity increase
Solution Approach 1:
The patent replaces manual mechanical interpretation processes with automated machine learning algorithms. The system uses trained neural networks to automatically identify geological layers and subsurface structures from seismic data, substituting human expert analysis with computational models that can process data rapidly without sacrificing accuracy.
Solution Approach 2:
The patent employs pre-trained machine learning models that have been previously trained on labeled seismic data. These pre-trained models can be directly applied to new seismic surveys without requiring manual interpretation, enabling rapid automated analysis while maintaining the accuracy that was achieved through extensive manual training and validation.
2Productivity
If automated machine learning methods are applied to seismic data analysis, then processing speed and productivity are improved, but accuracy and reliability of subsurface property inference deteriorate due to incomplete training data
Solution Approach 1:
The patent uses data augmentation techniques to create synthetic copies of limited seismic data. By generating artificial seismic surveys with known subsurface properties through physics-based forward modeling, the system expands the training dataset without requiring additional real-world seismic surveys, thereby improving model reliability while maintaining processing speed.
Solution Approach 2:
The patent incorporates multiple data dimensions and modalities into the training process, including seismic data from different surveys, well log data, and synthetic data with varying parameters. This multi-dimensional approach allows the machine learning model to learn more robust representations of subsurface properties, improving reliability without sacrificing the computational efficiency of automated processing.
3Loss of information
If comprehensive seismic surveys are conducted to obtain complete subsurface data, then data completeness and measurement precision are improved, but cost and operational complexity increase
Solution Approach 1:
The patent extracts and leverages useful information from existing, limited seismic surveys through advanced machine learning analysis. Instead of requiring comprehensive new surveys, the system extracts maximum subsurface property information from available data by using trained models that can infer properties not directly observable in the seismic records, thereby reducing the need for additional complex survey operations.
Data Source
AI summary
A method, apparatus, and program product utilize a disentangled factor learning framework to analyze petro-technical image data such as seismic image data to infer properties of a subsurface volume and/or to generate image data for use in training machine learning algorithms for use in petro-technical applications.