The Vibrational Structure Theory covers approaches on solving the nuclear Schrödinger equation for the accurate prediction of rotation-vibration spectra. This can be performed, e.g. with the Molpro software package [1], where the nuclear wave function is approximated as a Hartree product of one-mode wave functions and optimized variationally. This Vibrational Self Consistent Field (VSCF) approach [2,3,4] can be correlation corrected by describing the nuclear wave function as a linear combination of Hartree products, i.e. configurations, with the VSCF wave function as a reference. The configuration space dramatically increases with number of atoms, hence, in order to perform Vibrational Configuration Interaction (VCI) calculations the space has to be truncated and configuration-selection schemes have to be considered [5]. Nevertheless, for molecules with more then 10 atoms, the VCI calculation is one bottleneck. In some cases it needs hundreds of GB of memory for storing VCI subspaces during configuration-selection. Thus, only an architecture as provided by the MACH2, with its large amount of memory, allows for VCI calculations of bigger molecular systems.
Previous to the VSCF/VCI calculations, a Potential Energy Surface (PES) has to be computed at high level of Electronic Structure Theory, such as coupled cluster theory with perturbative triples (CCSD(T)) and large basis-sets. Although the PES generation algorithm reduces the number of ab inito electronic energy calculations (single points) by iterative interpolation techniques and symmetry considerations [6], the amount of single points increases steep with the number of atoms. Hence, the computation time of those CCSD(T) single points have to be kept in the range of some minutes. This can be established by the use of explicitly-correlation and orbital localization approximations. For extended molecular systems, the desired scaling of computation time with molecular size for such a single points can be reached, when e.g. 100 cores in combination with 1 TB of memory are available. When we perform the PES generation with 16 processes, we use 1600 cores with a total of 16 TB of memory. The MACH2 architecture, hence, enables the extension of our studies to bigger molecular systems, which are not possible to calculate on other Austrian HPC architectures mainly by the limited availability of memory.
References