High-Speed Pseudo-Orthogonalization for the Car-Parrinello Method
Tomoo Aoyama, Takatoshi Higuchi, Umpei Nagashima
Abstract
Open-access reader
Tomoo Aoyama, Takatoshi Higuchi, Umpei Nagashima
Abstract
Open-access reader
An approximate orthogonalization method called "pseudo-orthogonalization" was derived here from the Schmidt orthogonalization in a partial space. Precision of the orthogonalization is decimal 2 digits when the partial space of 1/3 of the Schmidt orthogonalization is used. The speed-up ratio is 3.3 with 1000 vectors of 1500th dimension. Although the pseudo-orthogonalization is high-speed, the precision is inadequate for general numerical calculations. Two possible approaches to improve this precision were considered: (a) introducing a band structure in the vectors, and (b) restricting the sign of the vector elements. These approaches improved the precision by a factor of 2.1 and 1.5, respectively, without increasing the CPU time.To demonstrate the applicability of this high-speed pseudo-orthogonalization for large-scale calculations, we coded a complex-form of it into the SCF part of the Car-Parrinello method. In calculations of the total energy for bulk-silicon, the difference in total energy calculated using the pseudo-orthogonalization and that using the Schmidt orthogonalization was less than O(-5) [a.u.]. Based on this accuracy, this pseudo-orthogonalization can be used to speed-up the Car-Parrinello method when it is applied to large-scale calculations.
A significance statement is not available in the OpenAlex record.
A contribution statement is not available in the OpenAlex record.
Method details are not available in the OpenAlex metadata.
Findings are not separately available in the OpenAlex metadata.
Limitations are not available in the OpenAlex metadata.
Application details are not available in the OpenAlex metadata.
An approximate orthogonalization method called "pseudo-orthogonalization" was derived here from the Schmidt orthogonalization in a partial space. Precision of the orthogonalization is decimal 2 digits when the partial space of 1/3 of the Schmidt orthogonalization is used. The speed-up ratio is 3.3 with 1000 vectors of 1500th dimension. Although the pseudo-orthogonalization is high-speed, the precision is inadequate for general numerical calculations. Two possible approaches to improve this precision were considered: (a) introducing a band structure in the vectors, and (b) restricting the sign of the vector elements. These approaches improved the precision by a factor of 2.1 and 1.5, respectively, without increasing the CPU time.To demonstrate the applicability of this high-speed pseudo-orthogonalization for large-scale calculations, we coded a complex-form of it into the SCF part of the Car-Parrinello method. In calculations of the total energy for bulk-silicon, the difference in total energy calculated using the pseudo-orthogonalization and that using the Schmidt orthogonalization was less than O(-5) [a.u.]. Based on this accuracy, this pseudo-orthogonalization can be used to speed-up the Car-Parrinello method when it is applied to large-scale calculations.
Key concepts: Orthogonalization, Mathematics, Algorithm, Applied mathematics, Statistical physics, Physics