TLDR: A novel Geophysics-Informed Neural Network (GINN) is introduced for model-based seismic inversion. It simultaneously generates spatially varying Point Spread Functions (PSFs) and predicts acoustic impedance (IP), overcoming limitations of traditional 1D wavelets. By embedding geophysical modeling into a self-supervised deep learning pipeline, GINN produces more accurate and higher-resolution subsurface images, improving reservoir characterization and noise reduction in complex geological settings.
Understanding the Earth’s subsurface is crucial for various applications, especially in the oil and gas industry for reservoir characterization. A key technique used for this is model-based seismic inversion, which helps estimate properties like porosity by analyzing seismic data. Traditionally, this process relies on simplified models of seismic waves, often using a “1D stationary average wavelet.” While useful, this simplification can lead to less accurate results because it doesn’t account for the complex, real-world variations in seismic data caused by how data is collected and the intricate geology underground.
An improvement over these 1D wavelets is the use of Point Spread Functions (PSFs). Think of a PSF as the unique “fingerprint” of how seismic energy travels and is recorded at a specific location. PSFs are spatially varying, meaning they change from one point to another, and they can capture the non-stationary behavior of seismic waves, providing a much more realistic picture of the subsurface, especially in areas with complex geological features. However, generating a full volume of PSFs is computationally very expensive and time-consuming, requiring specialized processing.
Introducing the Geophysics-Informed Neural Network (GINN)
To overcome these challenges, researchers have proposed a novel approach: the Geophysics-Informed Neural Network (GINN). This innovative deep learning architecture uses a Deep Convolutional Neural Network (DCNN) to simultaneously achieve two critical goals: generate a surrogate for the PSF and predict the acoustic impedance (IP). Acoustic impedance is a fundamental property that helps characterize subsurface rock formations.
The GINN works by embedding a 2D convolutional seismic modeling process directly into its pipeline. Here’s how it generally functions: the network takes real seismic amplitude data as input. It then generates estimates for reflection coefficients (RC) and PSFs. These generated RCs are then convolved (a mathematical operation similar to blending) with the generated PSFs to simulate seismic amplitude data. This “modeled” seismic amplitude is then compared to the original “true” seismic amplitude. The difference between the two forms a “loss function,” and by minimizing this error through a process called backpropagation, the GINN iteratively refines its internal weights to produce better estimates of both RCs and PSFs that closely match the observed seismic data.
Unlike some other physics-informed neural networks that embed explicit physical laws directly into their loss functions, the GINN incorporates convolutional seismic modeling processes that are deeply rooted in geophysical principles. This tailored integration of seismic inversion principles and deep learning techniques is why it’s termed a Geophysics-Informed Neural Network.
How the GINN Was Developed and Trained
For their experiments, the researchers used synthetic data derived from the SEAM Phase I Earth Model, a widely recognized benchmark for seismic inversion due to its realistic representation of complex subsurface structures like salt bodies and faults. The input to the GINN included Reverse Time Migration (RTM) seismic amplitude images and additional positional information for each pixel, helping the network account for variations related to depth and lateral position.
A crucial aspect of the training involved incorporating low-frequency impedance (LF-IP) information. Seismic data naturally lacks low-frequency content, which is vital for reconstructing absolute impedance contrasts. By including LF-IP, the GINN can produce more complete and accurate impedance estimations.
The network itself was built using a 2D UNet architecture, known for its effectiveness in image-to-image transformations. The GINN produces three outputs: pseudo-impedance (pseudo-IP), a proto zero-phase PSF, and a residual PSF. To ensure the generated PSFs are geophysically consistent, a regularization strategy was employed. This involved enforcing a “zero-phase” assumption for the PSFs, which is a common geophysical characteristic. The final PSF is a combination of a low-frequency zero-phase component and a high-frequency residual component, balancing adherence to principles with capturing fine details.
The training process involved comparing the modeled seismic amplitude with the true seismic amplitude using two loss functions: Mean Squared Error (MSE) for pixel-wise differences and Structural Similarity Index Measure (SSIM) for structural feature comparison. The model was trained for about 100 epochs, taking approximately 20 minutes.
Also Read:
- QUTCC: Enhancing Deep Learning Reliability in Imaging Inverse Problems
- A New Approach to Understanding Time Series Signals: Focusing on Structure Over Amplitude
Promising Results and Future Outlook
The results from the GINN predictions are highly promising. The generated impedance (IP) and reflection coefficient (RC) sections appeared continuous and provided a realistic geological representation of the subsurface. Notably, the resolution of these sections was significantly enhanced compared to the raw seismic amplitude data. The predicted PSFs also looked realistic, aligning with expected geological illumination and demonstrating the network’s ability to model spatially variable seismic resolution.
A key advantage highlighted is that the estimated PSFs exhibit a limited lateral resolution, which effectively reduces noise and improves the accuracy of the inversion process. This is a significant improvement over traditional 1D wavelets, which assume an unrealistic lateral resolution. This makes the GINN particularly robust for complex geological environments.
While the results are encouraging, the researchers noted some minor vertical artifacts in the impedance section, indicating areas for further investigation. Future work could focus on refining the training process, incorporating additional prior information, or exploring alternative loss functions to better constrain the solution space and minimize these artifacts. Validation with real seismic data would also be a crucial next step to assess the method’s applicability in practical scenarios.
In conclusion, the GINN pipeline represents a significant step forward in seismic inversion, integrating deep learning with geophysical principles to automatically estimate non-stationary PSFs and produce more accurate acoustic impedance approximations. This approach promises a more detailed and reliable understanding of the Earth’s subsurface. For more technical details, you can refer to the original research paper here.
