- Subsurface Velocity Model: Why It Matters
- Seismic Imaging Meets AI
- Convolutional Neural Network (CNN) Can See the Patterns
- Common-Image Gathers (CIG): What They Show
- Full-Waveform Inversion (FWI) vs CNN: A Contrast
- Migration Velocity Analysis (MVA): The Precursor
- Seismic Data Processing with Deep Learning
- Residual Moveout Becomes a Clue
- Synthetic Velocity Models: Training's Building Blocks
- Seismic Migration Quality Check via CNN
- Frequency-Domain Reverse Time Migration (RTM) Friendly
- Angle-Domain Common Image Gathers (ADCIG) & Offset-Domain CIG (ODCIG)
- Feature Extraction for the Win
- Training Data Set and Neural Network Training
- Seismic Events Guiding Velocity
- Computational Seismology Meets AI
- Synthetic and Field Data Application: What's Next?
- Deep Learning for Velocity Model Building with CIG Volumes: Step-by-Step Guide
- Conclusion
Subsurface Velocity Model: Why It Matters
Creating a clear subsurface velocity model is like making a map of what’s under the Earth’s surface. This model helps in seismic imaging, a process of turning the echoes from seismic waves into images of what’s below the surface. Traditional methods, such as full-waveform inversion (FWI) and migration velocity analysis (MVA), are accurate but slow and sometimes unreliable. Because of this, deep learning is being used to speed up the process and reduce errors.
Seismic Imaging Meets AI
In seismic imaging, we send sound waves into the ground and measure the echoes. The echoes are then turned into pictures using seismic migration. These images are stacked into common-image gathers (CIG), which include different views such as angle-domain CIG (ADCIG) and offset-domain CIG (ODCIG). When the velocity model is correct, the seismic events in the images are flat; when the model is wrong, the events curve. This is where deep learning comes into play.
Convolutional Neural Network (CNN) Can See the Patterns
A convolutional neural network (CNN) is great at spotting patterns in complex data, such as the ones found in CIG volumes. Researchers have used synthetic CIGs paired with “true” synthetic velocity models to train a CNN. This allows the network to learn how seismic data relates to velocity models and to improve the accuracy of the models.
Common-Image Gathers (CIG): What They Show
A CIG volume is created after seismic migration using a reference model. If the velocity model is wrong, the seismic events in the gather will appear curved. The CNN helps identify this residual moveout, or bending, and corrects it. It can detect large patterns, like the overall shape of the subsurface, and adjust the velocity model accordingly.
Full-Waveform Inversion (FWI) vs CNN: A Contrast
Full-waveform inversion (FWI) is a method used to adjust the velocity model to match the full wavefield of seismic data. While this method is very precise, it is also computationally heavy and can get stuck in incorrect solutions. On the other hand, the CNN learns from data and makes quick predictions after an initial training period, making it both fast and reliable.
Migration Velocity Analysis (MVA): The Precursor
Migration velocity analysis (MVA) is used to adjust the velocity model by looking at CIGs. While it’s effective, it requires manual steps and lots of computing power. With deep learning, the CNN can do what MVA does, but automatically and much faster.
Seismic Data Processing with Deep Learning
To train the CNN, a training data set is created:
- Generate many synthetic velocity models.
- Create corresponding CIG volumes.
- Train the CNN to find the relationship between the CIG curvature and velocity errors.
The neural network training process uses a combination of techniques to ensure both pixel accuracy and structural quality in the velocity models.
Residual Moveout Becomes a Clue
The CNN detects residual moveout, the way seismic events bend or curve, across the ray-parameter dimension. Then, the network uses this information to create a velocity model that reduces or eliminates the curvature.
Synthetic Velocity Models: Training’s Building Blocks
The network is trained on synthetic velocity models, which are constructed to represent a wide variety of subsurface scenarios. This helps the CNN learn general patterns, such as low wavenumber features that indicate large-scale velocity trends. These features are used to improve the model before refining it with more detailed adjustments.
Seismic Migration Quality Check via CNN
Once the CNN predicts a new velocity model, it is used in a seismic migration process. The quality of the resulting image is checked by looking at the migration image quality. The network also checks how well the seismic energy positioning is aligned with the predicted model. If the model is not accurate enough, the process can be adjusted using hybrid methods like FWI combined with CNN.
Frequency-Domain Reverse Time Migration (RTM) Friendly
When the velocity model is accurate, migrating seismic data using frequency-domain reverse time migration (RTM) produces high-quality images. Deep learning methods, such as those based on AG-ResUnet, often use the structural similarity index (SSIM) to ensure the seismic events remain properly aligned during migration.
Angle-Domain Common Image Gathers (ADCIG) & Offset-Domain CIG (ODCIG)
The CNN can work with both ADCIGs and ODCIGs, different types of CIGs, learning to detect errors and predict the correct velocities for each case.
Feature Extraction for the Win
One of the key benefits of CNNs is their ability to automatically extract features from seismic data. The CNN can identify edges, folds, and wavefronts in the data without needing explicit instructions from human operators. This makes seismic data processing much faster and more efficient.
Training Data Set and Neural Network Training
A strong training dataset is crucial for a reliable CNN. By training on a diverse set of synthetic models, the CNN learns to make accurate predictions for a wide range of subsurface environments. Neural network training ensures that the CNN becomes proficient at detecting and correcting velocity errors.
Seismic Events Guiding Velocity
The ultimate goal of seismic imaging is to get flat seismic events in CIGs. When the CNN predicts the correct velocity model, the seismic events are aligned and flattened. This serves as proof that the model is accurate and provides valuable insights into the subsurface.
Computational Seismology Meets AI
By combining seismic science with artificial intelligence, this approach is advancing the field of computational seismology. The initial effort in training the network pays off when the model is applied to real-world data, speeding up workflows in oil, gas, and earthquake studies.
Synthetic and Field Data Application: What’s Next?
The CNN performs well with synthetic data. As researchers continue to apply transfer learning to real-world datasets, this method will become even more useful in practical applications like oil and gas exploration and earthquake monitoring.
Deep Learning for Velocity Model Building with CIG Volumes: Step-by-Step Guide
- Build synthetic velocity models
- Generate CIG volumes using simulations
- Train the CNN on the CIG–velocity data pairs
- Test on synthetic data (validate the model)
- Apply the model to field data using transfer learning
- Migrate with CNN velocity model, check image quality
- Adjust if necessary using hybrid techniques like FWI + CNN
References
Conclusion
Deep learning for velocity model building with common-image gather volumes is revolutionizing seismic imaging and computational seismology. By training a CNN on synthetic velocity models and CIG volumes, this approach automates migration velocity analysis, speeds up model parameter estimation, and improves workflows for seismic migration. As AI continues to integrate with traditional methods like FWI and RTM, the future of velocity model building looks smarter, faster, and more reliable.
