Research in Domain Decomposition Methods for Eigenvalue Problems
This group is conducting two major research efforts under the general theme of parallel sparse matrix computations, plus a new project using data mining to study materials.
- New, novel algorithms for solving large eigenvalue problems - specifically for those related to electronic structure calculations. Current efforts emphasize Domain Decomposition technques on the one hand and spectrum slicing methods on the other. The researchers investigate parallel methods based on these two basic ideas to improve the runtime without sacrificing the accuracy of the computations. They will continue to develop eigenvalue solvers that will reduce the amount of time needed to compute eigenvalues of the Hamiltonian matrices. In their efforts, they will emphasize the development of eigensolvers that compute a very large number of eigenvalues and parallel eigensolvers. They also have a joint NSF-supported collaboration with Eric Polizzi (UMass Amherst), the developer of FEAST, which is a novel approach to the problem geared toward electronic structure calculations.
- Parallel robust iterative solvers. As three-dimensional models are gradually becoming commonplace, iterative methods for solving large sparse linear systems arising from the discretization of partial differential equations, are gaining popularity. In the past, direct methods have often been preferred for two-dimensional problems, especially on computers with large memories. However, there is currently a general consensus that iterative methods become almost mandatory for three-dimensional problems because of the challenging memory and computational requirements of these problems. With the use of iterative solvers comes the need for effective preconditioners. A new direction the group is now taking is to consider the use of low-rank approximation techniques for approximating the Schur complement systems. The researchers have recently developed this technique sequentially. A recursive version was implemented in MATLAB.The primary initial motivation of this research was the development of preconditioners for GPUs. However, the researchers are now expanding this viewpoint as they have realized that these preconditioners have excellent potential for a Domain Decomposition approach. The researchers now plan to develop these techniques for linear systems that arise from realistic applications such as computational fluid dynamics and other application areas. Special methods, such as eigenvalue deflation, will be used to enhance robustness.
- Materials informatics. The researchers have been working on using data mining-type methods for materials. Machine learning is a broad discipline that comprises a variety of techniques for extracting meaningful information and patterns from data. It draws on knowledge and "know-how" from various scientific areas such as statistics, graph theory, linear algebra, databases, mathematics, and computer science.Recently, materials scientists have begun to explore data-mining ideas for discovery in materials. This project will explore the power of these methods for studying various materials properties such as melting points, structures, formation energy, etc. The group also considers unsupervised learning such as clustering compounds based on data obtained, e.g. from the constituent atoms or band diagrams.
Return to this PI's main page.