Transfer Learning for Seismic Interpretation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ML-assisted seismic interpretation methods are prone to producing inaccurate predictions due to geophysical artifacts such as illumination shadows, 'migration smiles', and converted mode energy.
Innovation Solution
A method that utilizes a machine learning model trained with synthetic seismic data sets, including one set with labeled features and another set with injected noise based on a geological model, to identify features in field seismic data and generate a subsurface model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If ML models are trained using conventional methods without synthetic data augmentation, then the training process is simpler, but the model produces inaccurate predictions in the presence of geophysical artifacts
Solution Approach 1:
The patent creates synthetic seismic data sets that copy and simulate the characteristics of real seismic data, including geophysical artifacts. These synthetic data sets replicate the appearance and behavior of actual seismic signals, allowing the ML model to learn to distinguish between genuine geological features and artifacts without requiring extensive real data annotated with artifact labels.
Solution Approach 2:
The patent performs preliminary training of the ML model on synthetic seismic data sets before deploying it to interpret real field data. This preliminary action allows the model to pre-learn the patterns of geophysical artifacts and geological features in a controlled environment, so that when it encounters real data, it can accurately differentiate between the two without requiring complex real data preparation.
2Measurement precision
If synthetic seismic data sets with injected noise are used for training, then the model's ability to distinguish geological features from artifacts is improved, but the data generation and processing becomes more complex
Solution Approach 1:
The patent converts the harmful effect of noise and artifacts into a beneficial training mechanism. By injecting noise into synthetic seismic data sets during training, the patent teaches the ML model to recognize and filter out artifacts while maintaining the ability to identify genuine geological features. The noise that would normally degrade signal quality becomes a teaching tool that enhances the model's robustness and precision.
Solution Approach 2:
The patent systematically varies parameters such as noise level, artifact type, and data characteristics in the synthetic data sets during training. By changing these parameters across different training epochs and data sets, the model learns to handle a wide range of conditions and maintains high measurement precision across different seismic environments without requiring complex post-processing.
3Productivity
If transfer learning is applied using synthetic data, then the model can be trained more efficiently with less real data, but the initial model architecture and training pipeline becomes more complex
Solution Approach 1:
The patent creates a universal training pipeline that can handle both synthetic and real seismic data using the same ML model architecture and processing framework. The model is designed to be multi-functional, capable of learning from synthetic data for pre-training and then adapting to real field data without requiring separate specialized systems. This universality streamlines the overall process despite the added complexity of transfer learning.
Data Source
AI summary
A method includes receiving field seismic data that represents a subsurface, identifying features in the field seismic data using a machine learning model that was trained using at least one first synthetic seismic data set that includes one or more features and one or more labels of the features, and at least one second synthetic seismic data set, the first and second synthetic seismic data sets both generated based on a geological model. Noise is injected into the second synthetic seismic data based on the geological model. The method also includes generating a model of the subsurface based at least in part on the features that were identified in the field seismic data using the machine learning model.