Fixed-Point Optimization of Atoms and Density in DFT.

I describe an algorithm for simultaneous fixed-point optimization (mixing) of the density and atomic positions in Density Functional Theory calculations which is approximately twice as fast as conventional methods, is robust, and requires minimal to no user intervention or input. The underlying numerical algorithm differs from ones previously proposed in a number of aspects and is an autoadaptive hybrid of standard Broyden methods. To understand how the algorithm works in terms of the underlying quantum mechanics, the concept of algorithmic greed for different Broyden methods is introduced, leading to the conclusion that if a linear model holds that the first Broyden method is optimal, the second if a linear model is a poor approximation. How this relates to the algorithm is discussed in terms of electronic phase transitions during a self-consistent run which results in discontinuous changes in the Jacobian. This leads to the need for a nongreedy algorithm when the charge density crosses phase boundaries, as well as a greedy algorithm within a given phase. An ansatz for selecting the algorithm structure is introduced based upon requiring the extrapolated component of the curvature condition to have projected positive eigenvalues. The general convergence of the fixed-point methods is briefly discussed in terms of the dielectric response and elastic waves using known results for quasi-Newton methods. The analysis indicates that both should show sublinear dependence with system size, depending more upon the number of different chemical environments than upon the number of atoms, consistent with the performance of the algorithm and prior literature. This is followed by details of algorithm ranging from preconditioning to trust region control. A number of results are shown, finishing up with a discussion of some of the many open questions.