Robust linear-scaling optimization of compact localized orbitals in density functional theory

Abstract

Locality of compact one-electron orbitals expanded strictly in terms of local subsets of basis functions can be exploited in density functional theory (DFT) to achieve linear growth of computation time with systems size, crucial in large-scale simulations. However, despite advantages of compact orbitals the development of practical orbital-based linear-scaling DFT methods has long been hindered because a compact representation of the electronic ground state is difficult to find in a variational optimization procedure. In this work, we show that the slow and unstable optimization of compact orbitals originates from the nearly-invariant mixing of compact orbitals that are mostly but not completely localized within the same subsets of basis functions. We also construct an approximate Hessian that can be used to identify the problematic nearly-invariant modes and obviate the variational optimization along them without introducing significant errors into the computed energies. This enables us to create a linear-scaling DFT method with a low computational overhead that is demonstrated to be efficient and accurate in fixed-nuclei calculations and molecular dynamics simulations of semiconductors and insulators.