A low-computational-cost algorithm and its parallel implementation for periodic divide-and-conquer density-functional tight-binding (DC-DFTB) calculations are presented. The developed algorithm enables rapid computation of the interaction between atomic partial charges, which is the bottleneck for applications to large systems, by means of multipole- and interpolation-based approaches for long- and short-range contributions. The numerical errors of energy and forces with respect to the conventional Ewald-based technique can be under the control of the multipole expansion order, level of unit cell replication, and interpolation grid size. The parallel performance of four different evaluation schemes combining previous approaches and the proposed one are assessed using test calculations of a cubic water box on the K computer. The largest benchmark system consisted of 3,295,500 atoms. DC-DFTB energy and forces for this system were obtained in only a few minutes when the proposed algorithm was activated and parallelized over 16,000 nodes in the K computer. The high performance using a single node workstation was also confirmed. In addition to liquid water systems, the feasibility of the present method was examined by testing solid systems such as diamond form of carbon, face-centered cubic form of copper, and rock salt form of sodium chloride.
ASJC Scopus subject areas
- Computational Mathematics