Abstract
Monitoring Carbon Capture, Utilization, and Storage (CCUS) sites requires high-resolution subsurface imaging to track CO2 plume migration. Full Waveform Inversion (FWI) is the standard for this task but is traditionally computationally intensive. This study evaluates supervised and unsupervised deep learning (DL) approaches for FWI, specifically focusing on their performance when site-specific labeled data is limited. By comparing InversionNet, VelocityGAN, and Unsupervised Physics-based FWI (UPFWI), the research highlights the strengths of synthetic data augmentation and physics-informed loss functions in geophysical monitoring.
Introduction
Seismic Full Waveform Inversion (FWI) is a powerful technique used to convert recorded seismic waves into high-resolution velocity maps of the Earth's subsurface. In the context of CCUS, FWI is essential for monitoring the integrity of storage reservoirs and tracking the movement of injected CO2. However, traditional iterative FWI methods are computationally expensive. Deep learning offers a potential solution by providing near-instantaneous inversion once a model is trained. The primary hurdle for DL in geophysics is the "data hunger" problem: real-world sites rarely have the "ground truth" labels required for standard supervised training.
Methodology
The study compares three distinct deep learning architectures to address the challenges of subsurface imaging:
- InversionNet (Supervised): A convolutional neural network (CNN) based on an encoder-decoder architecture. It is trained to map seismic data directly to velocity models using a large labeled dataset.
- VelocityGAN (Supervised + Data Augmentation): This approach utilizes Generative Adversarial Networks (GANs) and "Style Transfer" techniques. By incorporating diverse textures from natural images into synthetic velocity models, the researchers created a robust training set that simulates complex geological features like folding and faulting.
- UPFWI (Unsupervised Physics-based FWI): This "label-free" method integrates the laws of physics directly into the training loop. The network predicts a velocity map, which is then validated using a forward modeling layer (the Acoustic Wave Equation). The loss is calculated by comparing the reconstructed seismic data with the original input, eliminating the need for ground truth velocity maps.
Datasets and Case Study
The models were evaluated using two primary data sources:
- OpenFWI: A benchmark dataset for seismic inversion. The "Style" family of this dataset was used to provide complex, non-geological textures that challenge the models' ability to generalize.
- San Juan Basin Geological Model: A site-specific model representing a real-world CCUS project. This included pre-injection background models and time-lapse reservoir simulations that track velocity changes caused by CO2 injection over 55 time steps.
Analysis and Insights
A key finding of the research is the effectiveness of the "Style" transfer approach. Training models on diverse, non-traditional textures prevents them from over-simplifying the subsurface as mere horizontal layers, leading to better performance on complex real-world structures.
Furthermore, the study demonstrates the utility of physics-informed constraints. While supervised models like InversionNet excel when data is abundant, UPFWI provides a path forward for sites where labeled data is non-existent. However, the unsupervised approach showed a higher tendency toward noise and overfitting, suggesting that physical constraints alone may not yet match the precision of well-trained supervised models.
Results
The evaluation focused on the models' ability to track the "Reservoir Velocity Change" over time.
- Supervised Performance: Both InversionNet and VelocityGAN, when trained on the "Style" dataset, accurately reconstructed the background geology and the evolving CO2 plume.
- Unsupervised Performance: UPFWI successfully matched physical method baselines and reconstructed seismic data, though the resulting velocity maps contained more artifacts than the supervised counterparts.
- Plume Tracking: The models successfully imaged the reservoir across critical time steps (40 through 55), clearly identifying the growth and migration of the CO2 plume within the San Juan Basin model.
Conclusion
This study underscores the trade-offs between supervised and unsupervised deep learning for geophysical applications. While supervised methods are highly accurate, their dependence on labeled data remains a bottleneck. Unsupervised physics-based methods offer a promising alternative but require further refinement to reduce noise. Future research will likely focus on hybrid frameworks that use synthetic data for pre-training and physics-based losses for site-specific fine-tuning, providing a robust toolset for long-term CCUS monitoring.
Summary updated