quantum-espresso/PW/examples/EXX_example/README

183 lines
10 KiB
Plaintext

Hybrid Hartree-Fock+DFT functionals are a still evolving feature in PWscf.
Some of the following notes may be obsolete.
WHICH FUNCTIONALS ARE IMPLEMENTED ?
The following hybrid functionals are implemented: Hartree-Fock, PBE0,
B3LYP, HSE (see Modules/functionals.f90 for updated info and more details).
The GAU-PBE functional (free from divergences at q->0) is also implemented.
Usually in PWscf the functional to be used is read from pseudopotential
files but we do not have so far a pseudopotential generator for hybrid
functionals so one needs to use pseudopotentials generated with some other
functionals (eg. LDA, PBE, BLYP) and force the usage of a hybrid functional
using input variable "input_dft" in system namelist; for instance,
input_dft="pbe0" will force the usage of PBE0 irrespective of the
functional used in the pseudopotential generation.
HOW DOES THE ALGORITHM WORK ?
The algorithm is quite standard: see for instance Chawla and Voth,
JCP {bf 108}, 4697 (1998); Sorouri, Foulkes and Hine, JCP {\bf 124},
064105 (2006); Spencer and Alavi, PRB {\bf 77}, 193110 (2008).
Basically, one generates auxiliary densities $\rho_{-q}=\phi^{*}_{k+q}*\psi_k$
in real space and transforms them to reciprocal space using FFT;
the Poisson equation is solved and the resulting potential is transformed
back to real space using FFT, then multiplied by $\phi_{k+q}$ and the
results are accumulated.
The only tricky point is the treatment of the $q\rightarrow 0$ limit,
which is described below, and in the Appendix A.5 of the QE paper
(note the reference to the Gygi and Baldereschi paper).
See J. Comp. Chem. {\bf 29}, 2098 (2008); JACS {\bf 129}, 10402 (2007)
for examples of applications.
HOW DOES SELF-CONSISTENCY WORK ?
The usage of hybrid functionals is VERY expensive (see later). Moreover
self-consistency should be reached on the density-matrix, instead of the
charge density as in traditional DFT. This is not feasible with plane waves.
The strategy used here is to consider an auxiliary set of wavefunctions psi
in addition to the usual set phi and to minimize the auxiliary functional
(let us focus on HF for simplicity):
E[phi,psi] = T[phi] + E_ext[phi] + E_Hartree[phi] +
<phi|Vx[psi]|phi> - 0.5*<psi|Vx[psi]|psi>
where Vx[psi] is the fock operator defined with the auxiliary function psi.
Taking the functional derivatives w.r.t. phi it can be shown that the scf
condition for phi are the HF equation with fixed Fock operator, so Vx does
not enter in the scf procedure and one can mix density as usual. The
minimum condition w.r.t. psi is simply psi=phi so when both psi and phi
are minimized the standard HF energy is obtained.
Actually one can show that the functional E[phi,psi] above is
E[phi,psi] = E_HF[phi] + dexx[phi,psi]
where dexx is a positive definite addition to E_HF .
The scf procedure goes as follow.
0) a normal scf (with LDA or similar functionals) is performed
1) hybrid functional is switched on and psi = phi (the current best wfcs)
2) a new scf is performed w.r.t phi, keeping fixed Vx[psi]
3) dexx[phi,psi] is computed and if it exceeds the required tolerance
the proceedure is repeated from point 1)
HF may require several phi-scf cycles to reach full convergence. B3LYP
and PBE0, due to the smaller fraction of HF exchange included, require
usually a smaller number of phi-scf cycles
HOW EXPENSIVE IS THE CALCULATION ?
Very expensive.
Applying the Fock operator on a single vawefunction (phi_k,v) requires
the calculation of an integral over the whole BZ and all psi bands.
For each needed pair psi_k+q,v' and phi_k,v an auxiliary charge density
rho(-q+G) is built in real space and then FFT to reciprocal space where
the corresponding Poisson equation is solved. This auxiliary potential is
FFT back in real space where it is multiplied by psi_k+q,v' and added to
Vx[psi]phi...
The cost of the operation is therefore roughly NBND * NQS * ( 2 * FFT + ... )
where NQS is the number of q-points chosen to represent the BZ integration,
and depends in general on the localization of the Wannier functions of the
system.
For comparison non-local pseudopotentials in the KB formulation (without
exploiting the locality of the KB projetors) cost NKB * (2 * NPW) where
NKB is typically of the order of NBND but NPW cost at least an order
of magnitude less than an FFT.
Therefore even when one can take NQS=1 (for large non-metallic system
this should be ok) hybrid functionals will require at least an order of
magnitude more resources that a standard calculation.
For Gamma-only calculations, some speedup can be obtained by computing
auxiliary densities $\rho_q$ on a "custom" FFT grid corresponding to a
cutoff (variable "ecutfock") smaller than the usual value 4*ecutwfc.
HOW CAN I CHOSE NQS IN INPUT ?
In the system namelist there are three variables nqx1,nqx2,nqx3 that
define the regular q-grid in the BZ in a way similar to the automatic
k-points generation. Their value must be compatible with the k-points
used (that is k+q must be equivalent to some other k in the k-points list)
Their default value are nqx1=1,nqx2=1,nqx3=1 (BZ integration is
approximated by gamma point value only).
DIVERGENCE AT q->0
The BZ integral to be performed has a diverging kernel when (q+G)->0
(except GAU-PBE, for which there is no divergent term: set
exxdiv_treatment='none', x_gamma_extrapolation=.false.)
This is dealt with by adding and subtracting a term with the same
divergence that can be integrated analytically and performing
numerically the integration for the non divergent residue
(variable exxdiv_treatment='gygi-baldereschi', default)
[Gygi-Baldereschi, PRB 34, 4405 (1986)]
One problem is left: the now non divergent q=0 term is not easily determined
since it is a 0/0 (non analytic) limit. Several options have been considered:
1) just discard it ... this is not a good idea in general because it
induces an error proportional to 1/(NQS*Omega) in the total energy
where Omega is the volume of the Wigner-Seitz cell of the crystal.
As one wish to keep NQS as small as possible this may be large.
2) exploit the fact that the term has the above dependence and extract
it from a calculation with a given nqx1,nqx2,nqx3 and the one with
a grid twice as coarse in each direction. One does not really need to
perform two calculations but can do it internally (even when nqx? are
not even numbers...). This seems to work and it is set as the default.
In order to disable this feature [and get back to option 1)] set
x_gamma_extrapolation = .false.
3) perform calculations in q-grids that are shifted away from gamma so that
the 0/0 term is not needed. This create some extra complication in the
coding and cannot be used with Gamma-only k-point integration.
In some tests it didn't seem superior to option 2) ... it was never
fully implemented and now it has been removed.
4) use the value at small (q+G) to estimate the (q+G)->0 limit. This
again has been tried and found to offer, for low order numerical
differentiation, no better results that option 2). It is possible
than higher-order formulas yield better results but this has not been
explored. This option is currently not implemented but it would be
easy to re-implement it.
5) use a spherical cutoff for coulomb potential (exxdiv_treatement='vcut_spheric')
In the case of strongly anisotropic supercells, such that one (or two) of the
products nki*ai (nki=number of k-points along axis i, ai=cell length along i axis)
is much larger than the others (for instance: nkz*az=50 A >> nky*ay=nkx*ax=10 A),
it can be shown that the fourier transform of 1/(q+G)**2 does not behave as 1/|r-r'|
for small q+G, thus producing instabilities with respect to the k point sampling.
In order to avoid this problem you have 2 possibilities:
1) change your supercell to a cubic one (all nki*ai of similar value)
2) use a real-space Wigner-Seitz cutoff. For this you have to turn on
exxdiv_treatment="vcut_ws" and converge your results with respect to
ecutvcut (reciprocal space cutoff for the correction, i.e.:
coulomb=1/(q+G)**2 for (q+G)**2> ecutvcut,
coulomb=cutoffed_coulomb for (q+G)**2 < ecutvcut).
Typical values for ecutvcut range from 0.7 to 2.0.
OTHER LIMITATIONS
So far NORM-CONSERVING and ULTRASOFT pseudopotentials are implemented.
The latter case requires the computation of density-like objects, in
principle defined on the complete FFT grid (cutoff "ecutrho > 4*ecutwfc).
In order to spare CPU time, this is done instead on the "smooth" FFT grid
(cutoff = 4*ecutwfc) in either G-space (very slow!) or in real space
(faster: option "tqr"). Note that "ecutfock" is not currently implemented
for US pseudopotentials. Hybrids with PAW are not currently functional.
PARALLEL IMPLEMENTATION ?
At present, both plane-wave and k-point parallelization have been
implemented. This is what is mostly needed for large systems.
An experimental parallelization on the band structure is also available
(pw.x -nbgrp N)
WHAT PROPERTIES CAN I COMPUTE ?
Energy and forces (thanks to Hellmann-Feynman theorem forces do not
require extra calculations). In principle also stresses but the
corresponding formulas have not yet been coded.
So structural optimization is OK if the cell shape is kept fixed.
Band structure ? yes and no. Obviously one computes wfc during the scf
cycle and their eigenvalues are printed in output.
This can be sufficient to draw a band structure or a DOS, but the problem
arises when one wishes non-scf calculations in k-points different from
those computed during the scf cycle. At present it is not possible
because this would require the knowledge of all bands at k+q that we do
not have. I do not know how to by-pass this problem.
ELECTRIC FIELD
I did not dig into this issue but Paolo Umari is using EXX with
electric field. For details it would be better to ask him directly.
AN EXAMPLE
run_example script in this directory performs two series of calculations:
1) total energy of Silicon using different values for nqx,
2) calculation of binding energy of o2,co,n2 from calculations in a
12 au cubic box and gamma sampling.
Running it will generate directory "results" to be compared with directory
"reference"
Please report problems and suggestions to the mailing list of QE developers:
<q-e-developers@qe-forge.org>