[Paper Review] A conjugate gradient method for electronic structure calculations

Xiaoying Dai, Zhuang Liu+2 more|arXiv (Cornell University)|Jan 28, 2016

Advanced Chemical Physics Studies19 references3 citations

TL;DR

This paper proposes a novel conjugate gradient (CG) method for electronic structure calculations within density functional theory, using a Hessian-based step size strategy and three orthogonality techniques (WY, QR, PD) to maintain orbital orthogonality. The CG method with QR-based orthogonality and Hessian-based step size outperforms existing methods like OptM-QR in convergence speed and stability, requiring fewer iterations and less computational time while achieving comparable accuracy.

ABSTRACT

In this paper, we study a conjugate gradient method for electronic structure calculations. We propose a Hessian based step size strategy, which together with three orthogonality approaches yields three algorithms for computing the ground state energy of atomic and molecular systems. Under some mild assumptions, we prove that our algorithms converge locally. It is shown by our numerical experiments that the conjugate gradient method is efficient.

Motivation & Objective

• To develop a more efficient and stable alternative to the self-consistent field (SCF) and gradient-type methods for Kohn-Sham density functional theory (DFT) calculations.
• To address the poor convergence and instability of SCF iterations, especially in large systems with small band gaps.
• To improve upon existing conjugate gradient methods by introducing a Hessian-based step size strategy and robust orthogonality maintenance via WY, QR, and polar decomposition (PD) approaches.
• To demonstrate that the proposed CG method converges locally under mild assumptions and outperforms state-of-the-art methods like OptM-QR in numerical experiments.

Proposed method

• Proposes a conjugate gradient method for minimizing the Kohn-Sham total energy functional under orthogonality constraints, using a Hessian-based step size derived from a local second-order Taylor expansion of the energy functional.
• Employs three orthogonality strategies—WY, QR, and polar decomposition (PD)—to maintain orthonormality of Kohn-Sham orbitals during iterations.
• Introduces an approximate Hessian (17) that neglects the Hartree and exchange-correlation terms, which are shown to contribute little to the Hessian, thus reducing computational cost while preserving accuracy.
• Uses a non-backtracking, Hessian-based initial step size, avoiding the expensive exact line search and the non-monotonic behavior of Barzilai-Borwein (BB) steps.
• Applies convergence analysis under an energy descent property, proving local convergence for all three algorithms (WY, QR, PD-based CG).
• Employs a modified conjugate direction update that combines the current gradient and previous search direction, ensuring descent and conjugacy.

Experimental results

Research questions

• RQ1 Can a conjugate gradient method with a Hessian-based step size strategy achieve faster convergence and better stability than existing gradient-type methods in electronic structure calculations?
• RQ2 How do different orthogonality maintenance techniques (WY, QR, PD) affect the performance and robustness of the conjugate gradient algorithm in Kohn-Sham DFT?
• RQ3 Is the approximate Hessian (17), which neglects the Hartree and exchange-correlation terms, sufficiently accurate for step size determination while reducing computational cost?
• RQ4 Does the Hessian-based step size strategy eliminate the need for backtracking line search, and how does it compare to BB step sizes in terms of convergence and stability?
• RQ5 How does the proposed CG method compare in efficiency and accuracy to the state-of-the-art OptM-QR algorithm with BB step sizes?

Key findings

• The QR-based conjugate gradient algorithm with Hessian-based step size (CG-QR-H) achieves the best performance, requiring the fewest iterations and least computational time across all tested systems.
• For benzene (C6H6), CG-QR-H reaches convergence in 119 iterations and 7.26 seconds, while OptM-QR-BB requires 164 iterations and 6.63 seconds, showing superior efficiency.
• In the large alanine chain (C33H11O11N11), CG-QR-H converges in 1232 iterations and 1204.82 seconds, whereas OptM-QR-BB requires 3190 iterations and 2338.59 seconds, demonstrating a significant speedup.
• The approximate Hessian (17) yields nearly identical convergence behavior to the exact Hessian (16), but with substantially reduced computational cost, making it the recommended choice.
• The Hessian-based step size avoids the non-monotonic behavior of BB steps, resulting in more stable convergence, especially in systems where OptM-QR shows instability.
• The algorithm using the approximate Hessian (17) is as accurate as the exact Hessian (16), and the Hartree and exchange-correlation terms contribute negligibly to the Hessian, validating the approximation.