Subsurface Elastic Models Using ML for Stable 4D Inversion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating elastic property data in subsurface volumes require months of laboratory measurements and fail to account for uncertainty, are computationally challenging, and do not incorporate machine learning in complex oil-water-gas systems, leading to unstable and noisy outputs.
Innovation Solution
Utilizing machine learning techniques, such as neural networks, to train models that generate elastic property data as a function of position and time, incorporating uncertainty from various sources, and applying these models to subsurface data measured in labs, fields, or simulations to derive quantities like shear-wave velocity and density changes efficiently and robustly.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If traditional laboratory measurement methods are used to generate elastic property data, then measurement precision may be maintained, but the process requires months of laboratory measurements and is computationally challenging
Solution Approach 1:
The patent replaces traditional mechanical laboratory measurement systems with machine learning models (neural networks) that process seismic data to generate elastic property data. The ML models are trained on laboratory measurement data to learn the relationships between seismic attributes and elastic properties, enabling rapid prediction without physical laboratory measurements.
Solution Approach 2:
The patent creates virtual copies of the laboratory measurement process through machine learning models. Instead of performing actual laboratory measurements on rock samples, the system uses trained ML models that replicate the measurement outcomes based on seismic data, significantly reducing time while maintaining accuracy through proper training.
2Reliability
If traditional inversion methods are used, then comprehensive elastic property data can be obtained, but the outputs are unstable and noisy
Solution Approach 1:
The patent replaces traditional iterative inversion algorithms with machine learning models that directly predict elastic properties from seismic data. This substitution eliminates the instability and noise inherent in traditional inversion methods by using the learned relationships from training data, providing more reliable and stable outputs.
Solution Approach 2:
The machine learning models perform self-optimization during the training process, automatically adjusting their internal parameters to minimize prediction errors. This self-service capability allows the models to learn robust relationships from training data and generalize well to new data, producing stable and reliable elastic property estimates without manual intervention.
3Device complexity
If machine learning techniques are applied to complex oil-water-gas systems, then computational requirements are reduced, but the models become more complex
Solution Approach 1:
The patent segments the complex oil-water-gas system into separate training datasets for different fluid scenarios. The machine learning models are trained independently on data from systems with different fluid compositions, allowing each model to specialize in specific conditions while maintaining overall system complexity manageable.
Solution Approach 2:
The patent uses parameter changes to handle different oil-water-gas systems by training models with varying input parameters representing different fluid properties. The models learn to adapt to different system conditions through parameter variations in the training data, reducing the need for separate complex models for each scenario.
Data Source
AI summary
Systems and methods are disclosed for generating elastic property data as a function of position and time. Exemplary implementations may include obtaining a first initial elastic model and a first set of elastic parameters, obtaining training subsurface data and a first training elastic property dataset, generating a first conditioned elastic model, and storing the first conditioned elastic model.