1: \begin{abstract}

2:   Fermi operator expansion (FOE) methods are powerful alternatives to

3:   diagonalization type methods for solving Kohn-Sham density functional

4:   theory (KSDFT). One example is the pole expansion and selected

5:   inversion (PEXSI) method, which approximates the Fermi operator by

6:   rational matrix functions and reduces the computational complexity to

7:   at most quadratic scaling for solving KSDFT. Unlike diagonalization

8:   type methods, the chemical potential often cannot be directly read off

9:   from the result of a single step of evaluation of the Fermi operator.

10:   Hence multiple evaluations are needed to be sequentially performed to

11:   compute the chemical potential to ensure the correct number of

12:   electrons within a given tolerance. This hinders the performance of

13:   FOE methods in practice.  In this paper we develop an efficient and

14:   robust strategy to determine the chemical potential in the context of

15:   the PEXSI method. The main idea of the new method is not to find the

16:   exact chemical potential at each self-consistent-field (SCF) iteration

17:   iteration, but to dynamically and rigorously update the upper and

18:   lower bounds for the true chemical potential, so that the chemical

19:   potential reaches its convergence along the SCF iteration. Instead of

20:   evaluating the Fermi operator for multiple times sequentially, our

21:   method uses a two-level strategy that evaluates the Fermi operators in

22:   parallel.  In the regime of full parallelization, the wall clock time

23:   of each SCF iteration is always close to the time for one single

24:   evaluation of the Fermi operator, even when the initial guess is far away from the converged

25:   solution.  We demonstrate the effectiveness of the new method using

26:   examples with metallic and insulating characters, as well as results

27:   from \textit{ab initio} molecular dynamics.  

28: \end{abstract}

29: