We will present new parallel implementations of a linear-scaling PNO-LMP2-F12 and PNO-LCCSD-F12 methods. The PNOs (pair natural orbitals) are generated by a sequence of transformations, using PAOs (projected atomic orbitals) and OSVs (orbital specific virtuals) in intermediate steps. This can be viewed as a sequence of tensor decomposition steps that lead to an enormeous reduction of the amount of data (amplitudes, integrals), without significant loss of accuracy. The transformed two-electron integrals are computed using local density-fitting methods. All integrals, amplitudes and other intermediate data are kept in distributed high-speed memory, thus avoiding any bottlenecks caused by disk I/O.  It will be demonstrated that the methods are scalable in the sense that the CPU-time and memory per processor become nearly independent of the molecular size if the number of processors used is proportional to the molecular size. The overall times are small as compared to the preceeding Hartree-Fock calculation, which means that the new methods are competetive with hybrid DFT methods. Benchmark calculations for reaction energies and reaction barriers will  be presented that demonstrate that near CCSD/CBS quality results can be obtained for large molecules. Extensions to open-shell and multi-reference cases will also be discussed and preliminary results be presented.

Scalable electron correlation methods

248th National Meeting (2014)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.