A trust-region augmented Hessian implementation for restricted and unrestricted Hartree-Fock and Kohn-Sham methods

TL;DR

This paper introduces a trust-region augmented Hessian method for Hartree-Fock and Kohn-Sham calculations that reliably converges, especially for challenging open-shell and antiferromagnetic molecules, often outperforming traditional DIIS methods.

Contribution

The paper presents a novel TRAH-SCF implementation that ensures convergence with fewer iterations and can find lower-energy symmetry-broken solutions compared to DIIS methods.

Findings

TRAH-SCF achieves convergence with tight thresholds in fewer iterations.

Often finds lower-energy symmetry-broken solutions than DIIS.

Maintains competitive runtime even with large basis sets.

Abstract

We present a trust-region augmented Hessian implementation (TRAH-SCF) for restricted and unrestricted Hartree-Fock and Kohn-Sham methods. With TRAH-SCF convergence can always be achieved with tight convergence thresholds, which requires just a modest number of iterations. Our convergence benchmark study and our illustrative applications focus on open-shell molecules, also antiferromagnetically coupled systems, for which it is notoriously complicated to converge the Roothaan-Hall self-consistent field (SCF) equations. We compare the number of TRAH iterations to reach convergence with those of Pulay's and Kolmar's (K) variant of the direct inversion of the iterative subspace (DIIS) method and also analyze the obtained SCF solutions. Often TRAH-SCF finds a symmetry-broken solution with a lower energy than DIIS and KDIIS. For unrestricted calculations, this is accompanied by a larger spin contamination, i.e. larger deviation from the desired spin-restricted <S^2> expectation value. However, there are also rare cases in which DIIS finds a solution with a lower energy than KDIIS and TRAH. In rare cases, both TRAH-SCF and KDIIS may also converge to an excited-state determinant solution. For those calculations with negative-gap TRAH solutions, standard DIIS always diverges. If comparable, TRAH usually needs more iterations to converge than DIIS and KDIIS because for every new set of orbitals the level-shifted Newton-Raphson equations are solved approximately and iteratively by means of an eigenvalue problem. Nevertheless, the total runtime of TRAH-SCF is still competitive with the DIIS-based approaches even if extended basis sets are employed, which is illustrated for a large hemocyanin model complex.