Online Adaptive Local-Global Model Reduction for Flows in Heterogeneous Porous Media

Yalchin Efendiev, Eduardo Gildin, Yanfang YangComput 2016 · DOI

TL;DR

This paper introduces an "online adaptive local-global POD-DEIM" method to speed up simulations of fluid flow in complex porous media. By intelligently updating a reduced-order model in real-time based on local error indicators, it achieves accurate results with significantly less computational cost, especially for scenarios where the system's behavior deviates from initial training data.

What Problem Does This Paper Address?

Simulating fluid flow in heterogeneous porous media (like oil reservoirs or groundwater systems) is computationally intensive. Traditional methods often rely on "reduced-order models" (ROMs) to speed things up. These ROMs are built offline using snapshots of the system's behavior. However, if the actual flow scenario during online simulation deviates significantly from the initial offline snapshots (e.g., due to changes in injection rates or unexpected geological features), the ROM's accuracy plummets. Existing adaptive methods often require expensive global updates or are limited in their ability to capture new dynamics. This paper tackles the challenge of creating an ROM that can adapt online to these deviations efficiently, using local information to guide global updates.

Methods

The core of the proposed method is an "online adaptive local-global POD-DEIM" framework. Here's a breakdown:

  • Offline Stage:
    • Proper Orthogonal Decomposition (POD): This standard technique is used to identify the dominant modes (basis functions) of the system's behavior from a set of "snapshots" (simulation results). This creates a low-dimensional subspace for the state variables.
  • Discrete Empirical Interpolation Method (DEIM): This method is used to efficiently approximate the nonlinear terms in the system equations, which are often a bottleneck in ROMs. It samples nonlinear functions at specific points and uses a reduced basis.
  • Generalized Multiscale Finite Element Method (GMsFEM): This is used to construct "local multiscale basis functions" which capture the behavior within specific regions of the porous media.
  • Online Stage (The Novelty):
    • Local Error Indicators: Instead of expensive global error calculations, the method uses local error indicators within the GMsFEM framework. These indicators highlight regions where the current ROM is performing poorly.
  • Local-Global Adaption: When local errors exceed a threshold, the system triggers an update. This update uses the local information to adapt the global POD subspace and DEIM approximations. This is done by solving a "residual problem" using GMsFEM, which is cheaper than solving the full-order system.
  • Adaptive POD: The POD subspace is updated by incorporating new basis functions derived from the residual problem. Two strategies are proposed: Adaptive-POD-1 (simply adding new modes) and Adaptive-POD-2 (selectively adding/removing modes based on their energy contribution).
  • Adaptive DEIM: The DEIM basis and interpolation points are also updated using efficient low-rank update techniques.

The key innovation is the local nature of the error detection and the online adaptation of the global model, guided by these local indicators.

Datasets

The paper uses numerical examples for:

  • Single-phase flow: A problem defined on a 2D domain with a heterogeneous permeability field. The "training" simulation uses specific injection rates, while the "test" simulation introduces new injection rates and a shut-off well, creating deviations from the training data.
  • Two-phase flow (oil-water): A 5-spot well configuration problem. The training simulation uses constant injection rates, while the test simulation introduces changes in well rates (shutting off/opening wells) to create deviations.

No new benchmark datasets are introduced; rather, existing problem setups are used to demonstrate the adaptive capabilities.

Key Insights and Ideas

  • Local Information for Global Adaptation: The most significant insight is that local error indicators, derived from multiscale basis functions, are sufficient to trigger effective global model updates. This bypasses the computational cost of global error estimation.
  • Efficient Online Updates: The method's ability to adapt the ROM during simulation, rather than relying solely on pre-computed offline data, is crucial for handling scenarios outside the initial snapshot space.
  • Robustness to Initial Snapshot Choice: The adaptive nature makes the method less sensitive to the initial choice of offline snapshots. Even if the initial snapshots don't perfectly capture the entire solution manifold, the online adaptation can absorb new information.
  • Balancing Accuracy and Cost: The infrequent nature of the global updates, guided by local indicators, ensures that the computational overhead of adaptation is manageable, leading to significant overall speedups.

Key Results

The paper demonstrates the effectiveness of the online adaptive local-global POD-DEIM method through numerical examples:

  • Single-phase flow: Achieved average L2 errors around 1% with adaptive POD, compared to ~18% without adaptation, even with a limited number of initial POD basis functions.
  • Two-phase flow:
    • For the pressure equation, adaptive POD significantly reduced saturation errors compared to static POD, with the most adaptive strategy yielding the best results.
  • For the saturation equation (which is nonlinear), POD-DEIM was applied. Adaptive-POD-2-DEIM showed better accuracy and used fewer basis functions than Adaptive-POD-1-DEIM.
  • For the coupled system, Adaptive-POD-2-DEIM again demonstrated superior performance in terms of accuracy and basis reduction compared to Adaptive-POD-1-DEIM.
  • In all cases, the adaptive methods showed a good agreement with the fine-grid reference solutions, especially when compared to non-adaptive approaches.

Limitations and Future Work

  • Future Work: The authors suggest finding more efficient ways to obtain the online basis for nonlinear systems. This implies that while the current method is efficient, further improvements in the online basis construction could yield even greater speedups.
  • Limitations (Implicit): The paper focuses on specific types of flow problems (single-phase and two-phase incompressible flow). While the methodology is presented as general, its application to more complex physics (e.g., compressible flow, multiphase flow with capillary effects, or different types of nonlinearities) might require further investigation and validation. The computational cost of the GMsFEM residual problem, while cheaper than the full-order system, could still be a factor for extremely large problems.

Summary updated

BibTeX
@article{efendiev2016online,
  author = {Yalchin Efendiev and Eduardo Gildin and Yanfang Yang},
  title = {{Online Adaptive Local-Global Model Reduction for Flows in Heterogeneous Porous Media}},
  journal = {comput.},
  year = {2016},
  doi = {10.3390/computation4020022},
}
Figure