TL;DR
This paper introduces a "theory-guided convolutional neural network" (TgCNN) framework to create accurate and efficient surrogate models for complex two-phase flow simulations in porous media. By integrating physical laws into the neural network training, TgCNN significantly improves prediction accuracy compared to purely data-driven models and enables faster, more reliable inverse modeling for reservoir characterization.
What Problem Does This Paper Address?
Simulating fluid flow in porous media, such as oil reservoirs or CO2 sequestration sites, is crucial for resource management and environmental studies. However, these simulations are computationally expensive, especially when performed repeatedly for inverse modeling (calibrating models to match observed data) or uncertainty quantification. Traditional methods struggle with the complexity of two-phase flow (e.g., oil and water interacting), which involves coupled governing equations and multiple state variables like pressure and saturation. Existing deep learning approaches, while promising, often ignore underlying physics, requiring vast amounts of training data and potentially leading to less robust predictions. This paper aims to bridge this gap by developing a physics-informed deep learning surrogate model that is both accurate and computationally efficient for two-phase flow problems.
Methods
The core innovation is the Theory-Guided Convolutional Neural Network (TgCNN). This framework extends previous work on single-phase flow to handle the more complex two-phase scenario. Key aspects include:
- Dual Network Architecture: Two separate convolutional neural networks (CNNs) are employed, one to approximate pressure and another for saturation. These networks are trained simultaneously.
- Physics-Informed Loss Function: The governing equations for two-phase flow are discretized and their residuals are incorporated directly into the neural network's loss function. This "theory-guidance" forces the network to adhere to physical principles, improving accuracy and reducing the need for extensive training data.
- Coupled Training: The two CNNs are mutually conditioned through the discretized governing equations, reflecting the inherent coupling between pressure and saturation in the physical system.
- Piecewise Training Strategy: For scenarios with changing operational conditions (like varying well controls over time), the TgCNN is trained in segments. Surrogates for each segment are then stacked to cover the entire simulation period, allowing for dynamic adaptation.
- Iterative Ensemble Smoother (IES) Integration: The trained TgCNN surrogates are integrated with the IES algorithm for efficient inverse modeling. This replaces computationally expensive simulator runs with fast surrogate predictions, significantly speeding up the parameter estimation process (e.g., inferring permeability fields).
Datasets
The study uses simulated data generated by the Eclipse reservoir simulator.
- Training Data: 80 permeability field realizations were generated using the Karhunen-Loeve expansion (KLE) and solved with Eclipse to create labeled datasets for training.
- Virtual Realizations: An additional 200 (for standard cases) or 800 (for increased heterogeneity) virtual realizations were used to impose physical constraints (discretized governing equations) during training.
- Testing Data: 200 test realizations were generated and solved with Eclipse to evaluate the accuracy of the trained TgCNN surrogates.
Key Insights and Ideas
- Physics is Key for Complex Flows: For intricate problems like two-phase flow, simply feeding data to a neural network isn't enough. Incorporating physical laws (theory-guidance) is crucial for achieving high accuracy and robustness, especially when data is limited.
- Coupled Networks Learn Interdependencies: Training separate networks for pressure and saturation simultaneously, and linking them through the governing equations, effectively captures their complex interplay, leading to better overall predictions.
- Piecewise Training for Dynamic Scenarios: The ability to train surrogates in segments for changing operational conditions (like well controls) is a significant advantage for real-world reservoir management, allowing models to adapt without complete retraining.
- TgCNN as a Powerful Inverse Modeling Tool: Combining TgCNN with IES offers a substantial speedup for inverse modeling, making it feasible to tackle more complex reservoir characterization problems that were previously computationally prohibitive.
- Robustness to Parameter Variance: The TgCNN-based IES shows remarkable robustness, performing well even when the variance of the permeability field in test cases differs from the training data, suggesting its applicability across a range of reservoir heterogeneities.
Key Results
- Improved Accuracy: TgCNN surrogates achieved significantly higher accuracy than purely data-driven CNNs for both pressure and saturation predictions, as evidenced by lower relative L₂ errors and higher R² scores. For instance, in Figure 6, TgCNN consistently shows lower L₂ errors and higher R² scores across pressure and saturation.
- Efficient Inverse Modeling: The TgCNN-based IES significantly reduced the computational cost of inverse modeling. For example, training the surrogate for a new variance scenario took approximately 1.864 hours with transfer learning, compared to over 9 hours for direct retraining.
- Accurate Production Predictions: The trained TgCNN surrogates, when used with IES, produced accurate matches for historical production and injection data and provided reliable predictions for future reservoir states. Figure 17 and 19 show good agreement between predicted and reference oil and water production/injection rates.
- Reduced Uncertainty: Inverse modeling with TgCNN-based IES effectively reduced the uncertainty in permeability field estimations, particularly around well locations, as shown in Figure 13.
- Robustness to Heterogeneity: The method demonstrated good performance even when the variance of the permeability field in the test cases differed from the training data, indicating its practical utility.
Limitations and Future Work
- Computational Cost for Training: While the TgCNN framework significantly speeds up inference and inverse modeling, the initial training process, especially for complex scenarios or with increased heterogeneity, can still be computationally intensive.
- Sensitivity to Variance (Not Infinite): Although the method is robust to variations in permeability field variance, this insensitivity is not absolute. The paper notes that performance will diminish as the real variance deviates further from the training data's variance.
- Future Work: The paper suggests that with the development of oil reservoirs, production data from subsequent periods can be collected to further refine inversion results, implying ongoing model improvement and adaptation. The authors also highlight the potential for using transfer learning to accelerate convergence for new scenarios, which is a key avenue for future efficiency gains.
Summary updated