Deep-learning-based Fault Delineation Using Passive Seismic Data: A Case Study at the Decatur CO2 Storage Site

Core Concepts

This research paper introduces DeFault, a novel deep learning method for accurately and efficiently relocating passive seismic events and delineating faults, showcasing its potential for ensuring the safety and success of CCUS projects by improving subsurface characterization.

Abstract

• Bibliographic Information: Wang, H., Chen, Y., Alkhalifah, T., Chen, T., Lin, Y., Alumbaugh, D. (2024). DeFault: Deep-learning-based Fault Delineation Using the IBDP Passive Seismic Data at the Decatur CO2 Storage Site. arXiv:2311.04361v3 [physics.geo-ph].
• Research Objective: This study aims to develop and validate DeFault, a machine learning-based fault detection method, using passive seismic monitoring data from the Decatur CO2 storage site to improve the efficiency and accuracy of fault detection and analysis in CCUS projects.
• Methodology: The researchers developed DeFault, a deep learning workflow, to analyze passive seismic data from the Decatur CO2 storage site. This involved:
  • Signal processing and enhancement: Applying filters to raw seismic data to remove noise and enhance effective signals.
  • Deep-learning-based event location: Training a neural network with labeled synthetic data generated using a 3D velocity model and applying it to field data using a domain adaptation technique called MLReal to bridge the gap between synthetic and real data distributions.
  • K-means clustering: Clustering the predicted microseismic event locations within specific timeframes using the K-means algorithm to identify potential fault lines.
  • Dropout uncertainty analysis: Implementing dropout layers in the neural network to quantify the uncertainty associated with event location predictions.
• Key Findings:
  • DeFault accurately relocated passive seismic events and delineated faults, showing consistent results with previous studies using modified double-difference methods.
  • The identified fault lines predominantly oriented in a north-east to south-west direction, aligning with previous observations of microseismic clusters at the Decatur site.
  • Short-term fracture delineation suggested that regional faults and fractures are reactivated multiple times during CO2 injection.
  • Uncertainty analysis revealed higher uncertainty for events in close proximity due to similar features, highlighting the need for data from additional monitoring stations.
• Main Conclusions:
  • DeFault offers a promising approach for efficient and accurate passive seismic event relocation and fault delineation, contributing to safer and more successful CCUS projects.
  • The study highlights the importance of domain adaptation techniques like MLReal in bridging the gap between synthetic training data and real-world field data.
  • The findings emphasize the need for continuous monitoring and analysis of induced seismicity in CCUS projects to ensure long-term storage integrity and minimize potential risks.
• Significance: This research significantly contributes to the advancement of CCUS technology by providing a novel deep learning method for improved subsurface characterization, potentially leading to safer and more efficient CO2 storage operations.
• Limitations and Future Research:
  • The study acknowledges limitations regarding the assumption of vertical axis independence in MLReal transformations and the reliance on the accuracy of the velocity model.
  • Future research should explore more advanced domain adaptation techniques and incorporate elastic effects in seismic simulations for enhanced accuracy.
  • Integrating data from additional monitoring stations could reduce uncertainty in event location predictions and improve fault delineation.

Stats

The researchers used data from the Illinois Basin-Decatur Project (IBDP), which included recordings from 31 sensors.
They synthetically generated a training set of 15,000 samples using a 3D interval velocity model derived from stacking velocity measurements.
The synthetic source locations for training were constrained to a depth range of [1.7, 2.4] km.
The median hypocentral difference between the catalog and DeFault results was 0.04229 km, while the mean difference was 0.06228 km.
The average uncertainty range for well-separated events was within 0.15 km, while events in close proximity had uncertainty up to 0.8 km.

Quotes

"In this study, we aim to develop and validate an ML-based fault detection method, DeFault, using passive seismic monitoring data from the Decatur CO2 site."
"Our approach accurately and efficiently relocated passive seismic events, identified faults and could aid in potential damage induced by seismicity."
"Our results highlight the potential of DeFault as a valuable tool for passive seismic monitoring, emphasizing its role in ensuring CCUS project safety."

Key Insights Distilled From

DeFault: Deep-learning-based Fault Delineation Using the IBDP Passive Seismic Data at the Decatur CO2 Storage Site

by Hanchen Wang... at arxiv.org 11-06-2024

https://arxiv.org/pdf/2311.04361.pdf

Deeper Inquiries

How might DeFault be adapted for use in other geological settings and for different types of subsurface fluid injection projects beyond CO2 storage?

DeFault, as a deep-learning-based fault delineation method, demonstrates significant potential for adaptation to various geological settings and subsurface fluid injection projects beyond CO2 storage. Here's how:

Transfer Learning and Domain Adaptation: The core principles of DeFault, particularly the MLReal domain adaptation technique, can be extended to other scenarios. By utilizing transfer learning, where a model trained on one task is used as a starting point for a model on a second, related task, DeFault can be fine-tuned with data from new geological settings. This minimizes the need for extensive training data from scratch. For instance, if applied to geothermal energy extraction, where induced seismicity is also a concern, a DeFault model pre-trained on CO2 storage data can be adapted using a smaller dataset specific to geothermal reservoirs.

Adjusting Training Data Generation: The synthetic training data generation process can be tailored to mimic the specific characteristics of different geological formations and injection scenarios. This includes incorporating:

Variable Velocity Models: Instead of relying solely on P-wave stacking velocity models, DeFault can integrate more complex 3D velocity models derived from seismic surveys, well logs, and geological interpretations. This allows for more accurate simulation of wave propagation in heterogeneous environments.
Different Source Mechanisms: While the current implementation uses a Ricker wavelet, the source function can be modified to represent the seismic signatures of other injection processes, such as wastewater disposal or enhanced oil recovery.
Incorporating Elastic Effects: Expanding the simulation to include elastic wave propagation, incorporating both P- and S-wave arrivals, would enhance the accuracy of event locations, especially in complex geological settings where S-wave anisotropy is significant.

Integrating Multi-Attribute Data: DeFault currently primarily utilizes seismic data. However, its architecture can be expanded to incorporate other geophysical data sources, such as:

Microgravity Data: Changes in fluid distribution due to injection can be reflected in microgravity measurements. Integrating this data could provide additional constraints on fluid migration pathways and fault activation.
Satellite Geodesy: Incorporating surface deformation data from Interferometric Synthetic Aperture Radar (InSAR) can help constrain fault geometry and slip behavior at shallower depths.

Application to Other Injection Projects: Beyond CO2 storage, DeFault's adaptability makes it suitable for monitoring a range of subsurface fluid injection projects, including:

Geothermal Energy Extraction: Monitoring induced seismicity during geothermal fluid injection is crucial for sustainable energy production.
Wastewater Disposal: Understanding fault activation related to wastewater injection can mitigate the risk of induced earthquakes.
Enhanced Oil Recovery: Monitoring microseismicity during fluid injection for enhanced oil recovery can optimize production and minimize environmental risks.

Could the reliance on synthetic training data and the inherent uncertainties in velocity models limit the generalizability and reliability of DeFault in real-world applications with complex geological structures?

Yes, the reliance on synthetic training data and uncertainties in velocity models can potentially limit the generalizability and reliability of DeFault, especially in complex geological settings. Here's a breakdown of the challenges and potential mitigation strategies:
Challenges:

Synthetic Data Limitations: While MLReal attempts to bridge the gap, synthetic data may not fully encapsulate the complexities of real-world seismic wave propagation. Factors like:

Anisotropy: Variations in seismic wave velocity with direction are often not fully captured in simplified velocity models.
Attenuation: Energy loss as waves travel through the Earth is complex and challenging to simulate accurately.
Scattering: Wave scattering due to small-scale heterogeneities can significantly impact seismic recordings.

Velocity Model Uncertainties: Errors in velocity models directly translate to errors in event locations. In complex geological structures with significant lateral and vertical velocity variations, these uncertainties can be substantial.
Overfitting to Training Data: If the synthetic data does not adequately represent the variability of real-world scenarios, the model might overfit to the training data and perform poorly on unseen data.
Mitigation Strategies:

Improving Synthetic Data Realism:

High-Fidelity Simulations: Employing more sophisticated and computationally intensive wave propagation modeling techniques, such as full-waveform inversion (FWI) or reverse-time migration (RTM), can generate more realistic synthetic data.
Data Augmentation: Introducing realistic noise and variations into the synthetic data can improve the model's robustness to noise and uncertainty in real-world data.

Refining Velocity Models:

Joint Inversion: Combining multiple geophysical datasets, such as seismic, gravity, and well logs, in a joint inversion framework can improve velocity model accuracy.
Continuous Model Updating: Implementing techniques to update the velocity model iteratively as new data become available can help refine event locations over time.

Hybrid Approaches:

Combining Physics and Machine Learning: Integrating DeFault with physics-based methods, such as double-difference relocation techniques, can leverage the strengths of both approaches.
Ensemble Modeling: Utilizing an ensemble of DeFault models trained on different realizations of velocity models and synthetic datasets can provide a more robust estimate of uncertainty.

What ethical considerations and potential societal impacts should be considered when deploying AI-driven technologies like DeFault for monitoring and managing subsurface resources?

Deploying AI-driven technologies like DeFault for subsurface resource management necessitates careful consideration of ethical implications and potential societal impacts:

Transparency and Explainability: The "black box" nature of some AI models raises concerns about transparency and accountability. Ensuring that DeFault's decision-making process is explainable and understandable by stakeholders, including regulators, operators, and potentially affected communities, is crucial for building trust.

Data Bias and Fairness: AI models are susceptible to biases present in the training data. If the data used to train DeFault reflects historical biases in data collection or interpretation, it could lead to unfair or discriminatory outcomes, potentially impacting marginalized communities disproportionately.

Privacy and Data Security: Subsurface resource management often involves sensitive data, including proprietary information and personal data related to land ownership or resource rights. Ensuring robust data privacy and security protocols is paramount to prevent unauthorized access or misuse.

Job Displacement and Economic Impacts: The automation potential of AI technologies like DeFault raises concerns about job displacement in sectors related to subsurface monitoring and analysis. It's essential to consider retraining and reskilling programs for workers potentially affected by such technological advancements.

Public Perception and Trust: Open communication with the public about the benefits, limitations, and potential risks of AI-driven technologies in subsurface resource management is crucial for fostering public trust and acceptance.

Environmental Justice: The deployment of DeFault should be assessed for potential environmental justice implications. Ensuring that the benefits and risks associated with subsurface resource management are distributed equitably across all communities, regardless of socioeconomic status or geographic location, is essential.

Regulatory Frameworks and Oversight: Existing regulatory frameworks may need to be adapted to address the unique challenges posed by AI-driven technologies in subsurface resource management. Establishing clear guidelines for data governance, model validation, and performance monitoring is crucial for ensuring responsible development and deployment.