Accelerated training of deep learning surrogate models for surface displacement and flow, with application to MCMC-based history matching of CO2 storage operations

Overview

• This paper presents a method for accelerating the training of deep learning surrogate models for surface displacement and flow, with an application to MCMC-based history matching of CO2 storage operations.
• The key contributions include:
  • Developing a framework for efficiently training deep learning surrogate models for surface displacement and flow simulations.
  • Applying the surrogate models to speed up MCMC-based history matching of CO2 storage operations.
  • Demonstrating the effectiveness of the approach on a real-world case study.

Plain English Explanation

The paper discusses a technique to speed up the training of deep learning models that can approximate the behavior of complex physical systems. These "surrogate models" can be used to quickly predict things like surface displacement and fluid flow, which is important for analyzing and optimizing processes like CO2 storage underground.

The researchers developed a framework to efficiently train these deep learning surrogate models. They then applied the surrogate models to a real-world problem of history matching - adjusting the simulation model to match observed data - for a CO2 storage operation.

By using the fast surrogate models instead of the full, detailed simulations, the history matching process was accelerated significantly. This could lead to faster and more efficient analysis and optimization of CO2 storage and other complex physical systems.

Technical Explanation

The paper develops a framework for training deep learning surrogate models that can approximate the behavior of complex physical systems, such as surface displacement and fluid flow simulations. The key technical contributions include:

1. Surrogate Model Architecture: The authors propose a deep neural network architecture consisting of graph convolutional layers and long short-term memory (LSTM) layers to capture the spatial and temporal dynamics of the physical system.

2. Accelerated Training: To speed up the training of the surrogate models, the authors utilize techniques such as transfer learning and meta-learning, which allow the model to leverage knowledge from previously trained models.

3. Application to MCMC-based History Matching: The trained surrogate models are used to accelerate the Markov Chain Monte Carlo (MCMC) -based history matching process for CO2 storage operations. The surrogate models provide fast predictions of the system behavior, allowing for more efficient exploration of the parameter space.

4. Case Study Evaluation: The effectiveness of the approach is demonstrated on a real-world case study of CO2 storage, where the surrogate models are shown to provide accurate predictions while significantly speeding up the history matching process.

Critical Analysis

The paper presents a promising approach for accelerating the training of deep learning surrogate models and applying them to complex physical systems. However, some potential limitations and areas for further research include:

1. Generalization Capabilities: The authors mention that the surrogate models are trained and evaluated on a specific case study. Additional research is needed to understand how well the models generalize to other types of physical systems or scenarios.

2. Uncertainty Quantification: The paper focuses on improving the computational efficiency of the history matching process, but does not explicitly address the quantification of uncertainties in the surrogate model predictions. Incorporating uncertainty quantification could be an important area for future work.

3. Scalability to Larger Systems: The case study presented is relatively small-scale. Evaluating the performance of the approach on larger, more complex physical systems could provide valuable insights into the scalability of the method.

4. Comparison to Other Surrogate Modeling Techniques: While the paper compares the proposed approach to the full-scale simulations, a comparison to other surrogate modeling techniques, such as reduced-order models or Gaussian processes, could provide a more comprehensive evaluation.

Conclusion

This paper presents a novel framework for accelerating the training of deep learning surrogate models for complex physical systems, with a focus on applications to CO2 storage operations. The key contributions include the development of an efficient surrogate model architecture and the application of the models to speed up MCMC-based history matching. The results demonstrate the potential of this approach to enable faster and more efficient analysis and optimization of complex physical processes, which could have important implications for fields such as energy, climate, and environmental engineering.