The recently developed DMoI/COSMO model for calculation of the electronic and geometrical structure of molecules in solution has been further tested ana optimized for the parallel system CRAY T3D at SDSC. In the DMoI/COSMO model, the solute molecule is embedded into a cavitv surrounded by solvent, which is represented by a dielectric continuum. The polarization of the dielectric continuum by the charge distribution of the solute results in a charge distribution on the cavity surface. In the COSMO method, the surface charges are obtained directly from the electrostatic potential on the cavity surface. This represents the major computational advantage. Since the 'DMoI/COSMO energy is fully variational, accurate gradients with respect to the solute coordinates can be calculated. The DMoI/COSMO theory and applications have been presented 1] and published [2]. DMol1COSMO became a part of our software offering within the DMoI 3.0.0 release. This indicates that the program was thoroughly debugged and strict quality assurance protocols were followed. Further validation of the method was performed, in particular optimization of the van der Waals radii and the computation of the non-electrostatic contributions. The COSMO procedure was optimized to some degree as well by, e.g. employing a more efficient linear equation solver. The DMoI/COSMO program was optimized for the CRAY T3D platform [3] and it is avail-able at SDSC for academic NIH users, according to the terms of the NIH grant. A distributed memory model is used for running DMoI on concurrent processors. Communication between processors occurs via messages, using the Message Passing Interface [4]. There are two distinct phases of computation in DMol program. In short, there is a domain of the program which depends on the molecular numerical integration grid and another domaln which is governed by the orbital basis set. The first domain involves setup of the integration grid, numerical integration and density synthesis, whereas the orbital part involves various matrix operations and diagonalization of the Hamiltonian matrix. At present the grid dependent part of DMol is fully parallelized whereas diagonalization is still done in single processor mode. The performance of DMol was tested for the zeolite model cluster Si8O25Hl8. The grid-dependent part of the calculations scales linearly with the number of CRAY T3D processors (up to 64), whereas diagonalization (non parallel) takes about 3% of the total wall clock time for a single processor run. We are currently incorporating a parallel diagonalization routine in DMol. [l] ACS meeting, Anaheim, 1995, 6th International Conference on the Application of the DFT in Chemistry and Physics, Paris, France, 1995 [2] J. Andzeim, Ch. Koelmel, A. KIamt, J. Chem. Phys. 103 (1995) 9312-9320. [3] Y.S. Li, M.C. Wrinn, J.M. Newsam, M.P. Sears, J. Comp. Chem. 16 (1995) 226. H. Pritchard, E. Bierwagen, Y.S. Li, J. Andzelm, to be published [4] Message Passing Interface Forum. 'Document for a Standard Message-Passing Interface"""""""", Technical Report No. CS-93-214 (revised), University of Tennessee, April 1994.
Showing the most recent 10 out of 270 publications
