A divide-and-conquer method with approximate Fermi levels for parallel computations

This study proposes two efficient algorithms of the linear-scaling divide-and-conquer self-consistent field (DC-SCF) method for parallel computations. These algorithms minimize the amount of network communication required for determining the common Fermi level by adopting an approximate Fermi level. One algorithm adopts the quasi-converged Fermi level during DC-SCF iterations, while the other uses the quasi-converged electron numbers of individual subsystems. A numerical assessment demonstrates the high parallel efficiency for both methods without loss in accuracy.