Iterative Solution Methods for Reduced-Order Models of Parameterized Partial Differential Equations

Thumbnail Image


Publication or External Link





This dissertation considers efficient computational algorithms for solving parameterized discrete partial differential equations (PDEs) using techniques of reduced-order modeling. Parameterized equations of this type arise in numerous mathematical models. In some settings, e.g. sensitivity analysis, design optimization, and uncertainty quantification, it is necessary to compute discrete solutions of the PDEs at many parameter values. Accuracy considerations often lead to algebraic systems with many unknowns whose solution via traditional methods can be expensive. Reduced-order models use a reduced space to approximate the parameterized PDE, where the reduced space is of a significantly smaller dimension than that of the discrete PDE. Solving an approximation of the problem on the reduced space leads to reduction in cost, often with little loss of accuracy.

In the reduced basis method, an offline step finds an approximation of the solution space and an online step utilizes this approximation to solve a smaller reduced problem, which provides an accurate estimate of the solution. Traditionally, the reduced problem is solved using direct methods. However, the size of the reduced system needed to produce solutions of a given accuracy depends on the characteristics of the problem, and it may happen that the size is significantly smaller than that of the original discrete problem but large enough to make direct solution costly. In this scenario, it is more effective to use iterative methods to solve the reduced problem. To demonstrate this we construct preconditioners for the reduced-order models or construct well-conditioned reduced-order models. We demonstrate that by using iterative methods, reduced-order models of larger dimension can be effective.

There are several reasons that iterative methods are well suited to reduced- order modeling. In particular, we take advantage of the similarity of the realizations of parameterized systems, either by reusing preconditioners or by recycling Krylov vectors. These two approaches are shown to be effective when the underlying PDE is linear. For nonlinear problems, we utilize the discrete empirical interpolation method (DEIM) to cheaply evaluate the nonlinear components of the reduced model. The method identifies points in the PDE discretization necessary for representing the nonlinear component of the reduced model accurately. This approach incurs online computational costs that are independent of the spatial dimension of the discretized PDE. When this method is used to assemble the reduced model cheaply, iterative methods are shown to further improve efficiency in the online step.

Finally, when the traditional offline/online approach is ineffective for a given problem, reduced-order models can be used to accelerate the solution of the full model. We follow the solution model of Krylov subspace recycling methods for sequences of linear systems where the coefficient matrices vary. A Krylov subspace recycling method contains a reduced-order model and an iterative method that searches the space orthogonal to the reduced space. We once again use iterative solution techniques for the solution of the reduced models that arise in this context. In this case, the iterative methods converge quickly when the reduced basis is constructed to be naturally well conditioned.