Mike Minkoff
Mathematics and Computer Science Division
Argonne National Laboratory
Two Numerical Problems in Chemical Dynamics Simulation
Abstract
In this talk we describe two problems in chemical simulation. The first involved the
calculation of Potential Energy Surfaces -- a high-dimensional data-fitting problem which is
central to large-scale calculations of chemical reaction paths. We are exploring the use of
Interpolating Moving Least-Squares (IMLS) methods of high-dimension and present some of the
applications and results thus far. The second problem involves the calculation of
reaction-rate coefficients in an ab initio manner. Reaction rate coefficients can be
obtained from exact quantum cumulative reaction probabilities (CRP) by eigenvalue
calculations. We are using parallel software libraries to parallelize large-scale CRP
calculations. These calculations involve Lanczos methods to obtain the eigenvalues of the
probability operator. The evaluation of two linear solves of Greens functions are required
at each iteration. These, in turn, are solved iteratively via GMRES. Thus the matrix-vector
multiply is a key computational kernel. Results for the underlying matrix-vector operations
via tensor matrices are discussed in single-processor experiments up to ten dimensions.
|