PEXSI
 All Classes Namespaces Files Functions Variables Friends Pages
Solving Kohn-Sham density functional theory

PEXSI provides the following interface routines for electronic structure calculation based on the Kohn-Sham density functional theory. An example routine is given in driver_ksdft.c.

1) Estimate the range of chemical potential

Starting from a rough estimate of the chemical potential \(\mu\), [muMin0, muMax0], obtain a refined interval for the chemical potential [muMinInertia, muMaxInertia].

...
{
/* Setup the input matrix in distributed compressed sparse column (CSC) format */
...;
/* Step 1. Estimate the range of chemical potential */
nrows,
nnz,
nnzLocal,
numColLocal,
colptrLocal,
rowindLocal,
HnzvalLocal,
isSIdentity,
SnzvalLocal,
temperature,
numElectronExact,
muMin0,
muMax0,
numPole,
maxIter,
numElectronTolerance,
ordering,
npPerPole,
npSymbFact,
comm,
&muMinInertia,
&muMaxInertia,
&muLowerEdge,
&muUpperEdge,
&numIter,
&muList,
&numElectronList,
&info
);
...;
}

See PPEXSIInertiaCountInterface for detailed information of its usage.

Note
The main purpose of the step of estimating the range of chemical potential is to allow a wide range of initial guess of the chemical potential. When the guess of the chemical potential is already accurate, such as the case in consecutive steps of the self-consistent-field (SCF) iteration, the Newton iteration used in the next step will be efficient enough. In such case the step of estimating the range of chemical potential can be skipped.

2) Solve electronic structure problem

After obtaining the range of the chemical potential and a reasonably accurate guess for the chemical potential. The electronic structure problem for obtaining the selected elements of the Fermi operator (a.k.a. the single particle density matrix) can be computed using PEXSI. The chemical potential is refined at the same time using Newton's method. Besides the Fermi operator, the energy density matrix (for computing the force) and the free energy density matrix (for computing the Helmholtz free energy) can be computed at the same time.

Note
The iteration of the chemical potential is usually in the inner loop of an electronic structure calculation. The self-consistent field (SCF) iteration serves as an outer loop. Because it usually takes a few SCF steps to converge the electron density, it is not necessary to achieve very high accuracy in the inner loop, especially during the initial few SCF steps.
...
{
/* Step 2. Solve KSDFT using PEXSI */
nrows,
nnz,
nnzLocal,
numColLocal,
colptrLocal,
rowindLocal,
HnzvalLocal,
isSIdentity,
SnzvalLocal,
temperature,
numElectronExact,
muInertia,
muMinInertia,
muMaxInertia,
gap,
deltaE,
numPole,
muMaxIter,
PEXSINumElectronTolerance,
ordering,
npPerPole,
npSymbFact,
MPI_COMM_WORLD,
DMnzvalLocal,
EDMnzvalLocal,
FDMnzvalLocal,
&muPEXSI,
&numElectron,
&muMinPEXSI,
&muMaxPEXSI,
&muIter,
muList,
numElectronList,
numElectronDrvList,
&info );
...;
}

See PPEXSISolveInterface for detailed information of its usage.

3) Post processing (optional)

After the chemical potential has converged, post-processing steps can be performed. Currently PEXSI supports the computation of the density of states (zero temperature) via counting the negative inertia.

...
{
/* Step 3. Post processing */
/* Compute the density of states (DOS) via inertia counting (without
* including finite temperature effects) */
nrows,
nnz,
nnzLocal,
numColLocal,
colptrLocal,
rowindLocal,
HnzvalLocal,
isSIdentity,
SnzvalLocal,
temperature,
numElectronExact,
muInertia,
muMinInertia,
muMaxInertia,
gap,
deltaE,
numPole,
muMaxIter,
PEXSINumElectronTolerance,
ordering,
npPerPole,
npSymbFact,
MPI_COMM_WORLD,
DMnzvalLocal,
EDMnzvalLocal,
FDMnzvalLocal,
&muPEXSI,
&numElectron,
&muMinPEXSI,
&muMaxPEXSI,
&muIter,
muList,
numElectronList,
numElectronDrvList,
&info );
...;
}

See PPEXSIRawInertiaCountInterface for detailed information of its usage.

Besides the DOS, the spatially resolved local DOS can also be computed without computing eigenvalues or eigenvectors.

...
{
/* Step 3. Post processing */
/* Compute the local density of states (LDOS).
*
* Only the first pole group participates in the computation of the
* selected inversion for a single shift. */
nrows,
nnz,
nnzLocal,
numColLocal,
colptrLocal,
rowindLocal,
HnzvalLocal,
isSIdentity,
SnzvalLocal,
eta,
ordering,
npSymbFact,
readComm,
localDOSnzvalLocal,
&info);
...;
}

See PPEXSILocalDOSInterface for detailed information of its usage.

Note
When mpisize > npPerPole, PPEXSILocalDOSInterface should not be executed by all processors, but only by a subgroup of processors. For more information see driver_ksdft.c.