Abstract
Simulating fluid flow in complex porous media like oil reservoirs or aquifers is computationally demanding due to multi-scale heterogeneities and nonlinear physics. This paper presents a space-time nonlinear upscaling framework that combines a non-local multi-continuum (NLMC) approach with deep learning. By identifying multiple continua per coarse block and using convolutional neural networks (CNNs) to learn local interactions, the framework provides a computationally efficient coarse-grained model that preserves fine-scale accuracy even in the absence of scale separation.
Introduction
Subsurface fluid flow simulation is critical for managing energy resources and environmental protection. However, rock properties such as permeability vary across vast scales. Traditional upscaling methods, which average properties over coarse blocks, often fail when scale separation is absent, when high-contrast features like fractures exist, or when the underlying physics is highly nonlinear. This research addresses these challenges by developing a "coarse" model that captures essential fine-scale physics through a non-local multi-continuum lens.
Methodology: The NLMC Framework
The proposed Non-Linear Multi-Continuum (NLMC) framework operates through three integrated stages to bridge the gap between fine-grid accuracy and coarse-grid efficiency.
- Continua Identification: Instead of a single average value, multiple "continua" (representing distinct features like rock matrix or fracture networks) are identified for each coarse block using test functions that act as macroscopic variables.
- Local Downscaling & Machine Learning: To model the interactions between these continua, the system solves nonlinear problems in "oversampled" regions ($K^+$). Because these local solves are computationally expensive, CNNs are trained to act as surrogates, mapping macroscopic variables to fine-grid solutions (specifically, nonlinear transmissibilities).
- Global Coupling: The macroscopic variables are solved across the entire domain using a variational formulation, ensuring mass balance for each continuum at the coarse scale.
The use of oversampling is a critical "non-local" feature, ensuring that the influence of neighboring regions is accounted for, which is essential in heterogeneous media.
Experimental Evaluation
The framework was tested on synthetic 2D datasets representing fractured porous media, upscaling a 160x160 fine grid to a 10x10 coarse grid. Two primary scenarios were evaluated:
- Unsaturated Flow: A nonlinear problem where permeability is pressure-dependent.
- Two-Phase Flow: A complex transport problem involving saturation and pressure over 700 time steps.
CNNs were trained on a global dataset generated from fine-grid calculations, using an 80:20 split for training and validation.
Results and Discussion
The NLMC framework demonstrated significant improvements over traditional upscaling methods ($u^{UP}$):
- Accuracy in Nonlinear Flow: NLMC achieved a relative $L_2$ error of 2.155%, whereas the classic approach resulted in 11.773% error.
- Two-Phase Flow Performance: The pressure error for NLMC was remarkably low at 0.281%, compared to 14.063% for classic upscaling.
- Surrogate Precision: The CNNs showed high precision in learning nonlinear transmissibilities, with RMSE values ranging from 1.7% to 5.4%.
These results highlight that a single average per grid block is insufficient for complex media. By tracking multiple continua and using ML to capture local complexity, the NLMC framework maintains high fidelity.
Conclusion and Future Work
The NLMC framework provides a robust solution for upscaling nonlinear flow in heterogeneous media. While the current 2D implementation shows great promise, future work will focus on extending the method to 3D reservoir models. This will require more complex CNN architectures and strategies to reduce the computational overhead of generating training data. Despite the "offline" training cost, the framework offers a significant advantage for applications requiring numerous rapid simulations, such as uncertainty quantification.
Summary updated