TY - JOUR

T1 - Local Fitting of the Kohn-Sham Density in a Gaussian and Plane Waves Scheme for Large-Scale Density Functional Theory Simulations

AU - Golze, Dorothea

AU - Iannuzzi, Marcella

AU - Hutter, Jürg

PY - 2017/5/9

Y1 - 2017/5/9

N2 - A local resolution-of-the-identity (LRI) approach is introduced in combination with the Gaussian and plane waves (GPW) scheme to enable large-scale Kohn-Sham density functional theory calculations. In GPW, the computational bottleneck is typically the description of the total charge density on real-space grids. Introducing the LRI approximation, the linear scaling of the GPW approach with respect to system size is retained, while the prefactor for the grid operations is reduced. The density fitting is an O(N) scaling process implemented by approximating the atomic pair densities by an expansion in one-center fit functions. The computational cost for the grid-based operations becomes negligible in LRIGPW. The self-consistent field iteration is up to 30 times faster for periodic systems dependent on the symmetry of the simulation cell and on the density of grid points. However, due to the overhead introduced by the local density fitting, single point calculations and complete molecular dynamics steps, including the calculation of the forces, are effectively accelerated by up to a factor of ∼10. The accuracy of LRIGPW is assessed for different systems and properties, showing that total energies, reaction energies, intramolecular and intermolecular structure parameters are well reproduced. LRIGPW yields also high quality results for extended condensed phase systems such as liquid water, ice XV, and molecular crystals.

AB - A local resolution-of-the-identity (LRI) approach is introduced in combination with the Gaussian and plane waves (GPW) scheme to enable large-scale Kohn-Sham density functional theory calculations. In GPW, the computational bottleneck is typically the description of the total charge density on real-space grids. Introducing the LRI approximation, the linear scaling of the GPW approach with respect to system size is retained, while the prefactor for the grid operations is reduced. The density fitting is an O(N) scaling process implemented by approximating the atomic pair densities by an expansion in one-center fit functions. The computational cost for the grid-based operations becomes negligible in LRIGPW. The self-consistent field iteration is up to 30 times faster for periodic systems dependent on the symmetry of the simulation cell and on the density of grid points. However, due to the overhead introduced by the local density fitting, single point calculations and complete molecular dynamics steps, including the calculation of the forces, are effectively accelerated by up to a factor of ∼10. The accuracy of LRIGPW is assessed for different systems and properties, showing that total energies, reaction energies, intramolecular and intermolecular structure parameters are well reproduced. LRIGPW yields also high quality results for extended condensed phase systems such as liquid water, ice XV, and molecular crystals.

UR - http://www.scopus.com/inward/record.url?scp=85019112079&partnerID=8YFLogxK

U2 - 10.1021/acs.jctc.7b00148

DO - 10.1021/acs.jctc.7b00148

M3 - Article

AN - SCOPUS:85019112079

VL - 13

SP - 2202

EP - 2214

JO - Journal of Chemical Theory and Computation

JF - Journal of Chemical Theory and Computation

SN - 1549-9618

IS - 5

ER -