SIESTA-PEXSI: Massively parallel method for efficient and accurate \textit{ab initio} materials simulation without matrix diagonalization

TL;DR

This paper introduces SIESTA-PEXSI, a massively parallel, matrix diagonalization-free method for large-scale electronic-structure calculations within KSDFT, achieving high accuracy and scalability on high-performance computing systems.

Contribution

It combines the PEXSI technique with SIESTA to enable efficient, scalable, and accurate electronic structure calculations without eigenvalue computations.

Findings

Achieves comparable accuracy to traditional diagonalization methods.

Demonstrates high scalability on over 10,000 processors.

Successfully applies to large 1D, 2D, and bulk systems with various electronic properties.

Abstract

We describe a scheme for efficient large-scale electronic-structure calculations based on the combination of the pole expansion and selected inversion (PEXSI) technique with the SIESTA method, which uses numerical atomic orbitals within the Kohn-Sham density functional theory (KSDFT) framework. The PEXSI technique can efficiently utilize the sparsity pattern of the Hamiltonian and overlap matrices generated in SIESTA, and for large systems has a much lower computational complexity than that associated with the matrix diagonalization procedure. The PEXSI technique can be used to evaluate the electron density, free energy, atomic forces, density of states and local density of states without computing any eigenvalue or eigenvector of the Kohn-Sham Hamiltonian. It can achieve accuracy fully comparable to that obtained from a matrix diagonalization procedure for general systems, including metallic systems at low temperature. The PEXSI method is also highly scalable. With the recently developed massively parallel PEXSI technique, we can make efficient use of more than $10,000$ processors on high performance machines. We demonstrate the performance and accuracy of the SIESTA-PEXSI method using several examples of large scale electronic structure calculations, including 1D, 2D and bulk problems with insulating, semi-metallic, and metallic character.