Large-scale all-electron density functional theory calculations using an enriched finite element basis

TL;DR

This paper introduces an enriched finite element basis for large-scale all-electron density functional theory calculations, significantly improving efficiency and scalability while maintaining high accuracy across various systems.

Contribution

The authors develop a novel enriched finite element basis with atom-centered functions, enabling efficient large-scale DFT calculations with improved accuracy and parallel scalability.

Findings

Achieves ~1 mHa accuracy in ground-state energies.

Reduces computational time by 50-300 times compared to classical methods.

Demonstrates good parallel scalability up to 384 processors.

Abstract

We present a computationally efficient approach to perform large-scale all-electron density functional theory calculations by enriching the classical finite element basis with compactly supported atom-centered numerical basis functions that are constructed from the solution of the Kohn-Sham (KS) problem for single atoms. We term these numerical basis functions as enrichment functions, and the resultant basis as the enriched finite element basis. The enrichment functions are compactly supported through the use of smooth cutoff functions, which enhances the conditioning and maintains the locality of the basis. The integrals involved in the evaluation of the discrete KS Hamiltonian and overlap matrix in the enriched finite element basis are computed using an adaptive quadrature grid based on the characteristics of enrichment functions. Further, we propose an efficient scheme to invert the overlap matrix by using a block-wise matrix inversion in conjunction with special reduced-order quadrature rules to transform the discrete Kohn-Sham problem to a standard eigenvalue problem. Finally, we solve the resulting standard eigenvalue problem by using a Chebyshev polynomial based filtering technique to compute the relevant eigenspectrum. We demonstrate the accuracy, efficiency and parallel scalability of the proposed method on semiconducting and heavy-metallic systems of various sizes, with the largest system containing 8694 electrons. We obtain accuracies in the ground-state energies that are within $\sim 1$mHa with ground-state energies obtained from classical finite element as well as gaussian basis sets. We observe a $50-300$ and $\sim 8$ fold reduction in the overall computational time when compared to classical finite element and gaussian basis, respectively. We also observe good parallel scalability up to 384 processors for a benchmark system comprising of 280-atom silicon nano-cluster.