Howard C. Elman,Virginia Forstall
Abstract:Reduced-order modeling is an efficient approach for solving parameterized discrete partial differential equations when the solution is needed at many parameter values. An offline step approximates the solution space and an online step utilizes this approximation, the reduced basis, to solve a smaller reduced problem at significantly lower cost, producing an accurate estimate of the solution. For nonlinear problems, however, standard methods do not achieve the desired cost savings. Empirical interpolation methods represent a modification of this methodology used for cases of nonlinear operators or nonaffine parameter dependence. These methods identify points in the discretization necessary for representing the nonlinear component of the reduced model accurately, and they incur online computational costs that are independent of the spatial dimension $N$. We will show that empirical interpolation methods can be used to significantly reduce the costs of solving parameterized versions of the Navier-Stokes equations, and that iterative solution methods can be used in place of direct methods to further reduce the costs of solving the algebraic systems arising from reduced-order models.
What problem does this paper attempt to address?
The main problem that this paper attempts to solve is to improve the computational efficiency when solving nonlinear partial differential equations with parameters (especially the steady - state Navier - Stokes equations). Specifically, the author focuses on how to significantly reduce the computational cost required to solve these equations through reduced - order modeling (ROM) and empirical interpolation methods (EIM), especially when dealing with problems with nonlinear operators or non - affine parameter dependencies.
### Overview of the Main Problems
1. **Computational Cost of High - Fidelity Solutions**:
- For many practical applications (such as uncertainty quantification, design optimization, and sensitivity analysis), it is necessary to solve parameterized partial differential equations (PDEs) under a large number of different parameters. Traditional high - fidelity solution methods require high computational costs each time the parameter changes.
- To reduce this cost, reduced - order models are proposed to approximately solve by projecting the model onto a low - dimensional space.
2. **Challenges of Nonlinear Problems**:
- For linear and parameter - affine - dependent problems, standard reduced - order methods can significantly reduce the cost. However, for nonlinear problems or problems with non - affine parameter dependencies, the standard methods cannot achieve the expected cost savings.
- The reduced - order models of nonlinear problems still depend on the dimension \( N \) of the full model, resulting in high computational costs in the online stage.
3. **Application of Empirical Interpolation Methods**:
- The empirical interpolation method accurately represents the nonlinear components by interpolating the nonlinear part and selecting key points in the discretization grid, so that the computational cost in the online stage is independent of the dimension of the full model.
- By combining the empirical interpolation method and the iterative solution method, the cost of solving the algebraic system can be further reduced.
### Specific Problem Descriptions
- **Target Equations**: The paper focuses on solving the steady - state incompressible Navier - Stokes equations with random viscosity coefficients:
\[
-\nabla \cdot (\nu(\cdot, \xi) \nabla \vec{u}(\cdot, \xi)) + \vec{u}(\cdot, \xi) \cdot \nabla \vec{u}(\cdot, \xi) + \nabla p(\cdot, \xi) = f \quad \text{in } D \times \Gamma
\]
\[
\nabla \cdot \vec{u}(\cdot, \xi) = 0 \quad \text{in } D \times \Gamma
\]
\[
\vec{u}(\cdot, \xi) = b \quad \text{on } \partial D \times \Gamma
\]
where \(\vec{u}(\cdot, \xi)\) is the flow velocity, \(p(\cdot, \xi)\) is the scalar pressure, \(b\) determines the Dirichlet boundary condition, and the viscosity coefficient \(\nu(\cdot, \xi)>0\).
- **Construction of Reduced - Order Models**: The paper introduces how to use the discrete empirical interpolation method (DEIM) to construct reduced - order models and verifies their effectiveness and accuracy through numerical experiments.
By solving these problems, the paper aims to provide a new computational framework for efficiently solving complex nonlinear partial differential equations, especially suitable for scenarios where the same type of equation needs to be solved multiple times but with different parameters.