A primary goal of this project is to create framework for developing more robust and user-friendly solvers for linear systems of equations, which are ubiquitous in science, engineering, and industrial applications. The investigators will carry out integrated analysis and development of geometric multigrid and algebraic multigrid methodologies. Geometric multigrid (GMG) methods form a class of multilevel solvers designed to solve linear systems of equations arising from certain classes of discretized partial differential equations (PDEs). Algebraic multigrid (AMG) methods are also multilevel solvers, though these techniques avoid any dependence on information regarding an underlying grid geometry or PDE. As a result, GMG methods are effective tools for solving a more restricted class of linear systems with a strong theoretical backing for their performance, whereas AMG solvers apply to more general linear systems, though do not share the same mathematical rigor in justifying their performance. This research project will employ functional analysis as a natural framework for studying GMG and AMG methods in a unified setting. The resulting theory will establish strong guiding principles for analyzing and developing robust, efficient, and scalable multilevel solvers. The iterative solvers under development will be implemented in open source parallel codes, made available to a broader scientific computing community, providing powerful tools for simulation and a foundation for future algorithm research and development. Moreover, this project provides Ph.D. students with opportunities to participate in a variety of education and research activities, in which they will receive advanced training, participate in conferences, and collaborate with researchers from industry and Department of Energy laboratories.
This research project investigates a unifying framework for analyzing GMG and AMG methods in a functional analysis setting. This framework provides a strong foundation for understanding the operators, relationships between spaces, and other core ingredients involved in these multilevel solvers, such as the construction of coarse spaces and their respective bases, and a convergence analysis that applies to a broad variety of existing AMG methods. Furthermore, for problems originating from discretized PDEs, the integration of geometric information for constructing auxiliary grid-based preconditioners and highly effective smoothers on each level can be seamlessly integrated, allowing for more flexible and aggressive algebraic coarsening. More benefits of incorporating geometric information are realized in nearly optimal load balancing and predictable communication patterns for parallel implementations. For more general linear systems, the project studies the use of (relaxed) compressed sensing techniques to preserve sparsity for coarse space operators, which can be supplemented with information from an underlying geometric grid or adjacency graphs of the matrix to gain control over the computational complexity. It is expected that these techniques will be useful in solving problems with non-quasiuniform underlying grids or problems with matrices that are not symmetric and positive-definite (SPD). To verify the efficacy of the developed methodologies, the resulting solvers will be applied to fluid-structure interaction problems and nearly singular SPD problems.
