| 468
|
|
|
Jose M Soler |
7 years ago
|
 |
|
| 467
|
|
|
Jose M Soler |
7 years ago
|
|
|
| 466
|
|
|
Jose M Soler |
7 years ago
|
|
|
| 465
|
|
|
Jose M Soler |
9 years ago
|
|
|
| 464
|
|
|
Jose M Soler |
9 years ago
|
|
|
| 463
|
|
|
Jose M Soler |
9 years ago
|
|
|
| 462
|
|
|
Jose M. Soler |
9 years ago
|
|
|
| 461
|
|
|
Alberto Garcia |
9 years ago
|
|
|
| 460
|
|
|
Alberto Garcia |
9 years ago
|
|
|
| 459
|
|
|
Alberto Garcia |
10 years ago
|
|
|
| 458
|
|
|
Alberto Garcia |
10 years ago
|
|
|
| 457
|
|
|
Nick Papior Andersen |
10 years ago
|
|
|
| 456
|
|
|
Alberto Garcia |
10 years ago
|
|
|
| 455
|
|
Charge-density mixing
With the fdf option
MixCharge T
the self-consistent-field convergence acceleration will be carried out
by mixing the charge density (actually the Fourier-components of the
charge). Both a simple Kerker-preconditioning and a DIIS procedure with
a subset of the Fourier components are implemented.
The implementation follows the ideas in Kresse+Furthmuller,
Comp.Mat.Sci. 6, 15-50 (1996), KF in what follows).
Relvant parameters:
ThomasFermiK2 (physical energy)
determines the Thomas-Fermi wavevector squared (in Ry) featuring in
the Kerker mixing, which is always performed on all components of
rho(G), even those treated with the DIIS scheme. Default: 0 Ry.
RhoGMixingCutoff (physical energy)
determines the sub-set of G vectors which will undergo the DIIS
procedure. Only those with kinetic energies below this cutoff will be
considered. The optimal extrapolation of the rho(G) elements will be
replaced in the fourier series before performing the Kerker
mixing. Default: 9 Ry.
RhoG-DIIS-Depth
determines the maximum number of previous steps considered in the DIIS
procedure. Default: 0
RhoG-SP-Preconditioner-Cutoff (physical energy)
determines the value of "q1^2" in the weighing of the different G
components in the scalar products among residuals in the DIIS
procedure. Following the KF ansatz, this parameter is chosen so that
the smallest (non-zero) G has a weight 20 times larger than that of
the smallest G vector in the DIIS set. Default: The KF prescription.
DebugDIIS
controls the level of debugging output in the DIIS procedure. Default:
False.
DebugRhogMixing
controls the level of debugging output in the mixing procedure
(basically whether the first few stars worth of Fourier components are
printed). Default: False. Note that this feature will only display
the components in the master node...
SCF.MixCharge.SCF1
controls whether mixing is done in the first scf step. Default: false. (This
seems to speed up convergence in some cases. More data is needed.)
The information about rhog_in and rhog_out in first scf iteration is
not added to the DIIS history. This seems to improve the convergence.
* Added h2o-rhog-mixing test
Note that the default convergence criteria remains based on the
differences in the density matrix, but in this case the differences
are from step to step, not the more fundamental DM_out-DM_in. Perhaps
some other criterion should be made the default (max |Delta rho(G)|,
convergence of the free-energy...)
Note that with charge mixing the Harris energy as it is currently
computed in Siesta loses its meaning, since there is no "DM_in". The
program prints zeroes in the Harris energy field.
Note that the KS energy is correctly computed throughout the scf cycle,
as there is an extra step for the calculation of the charge stemming from
DM_out, which also updates the energies. Forces and final energies are
correctly computed with the final DM_out, regardless of the setting of
the option for mixing after scf convergence.
The program now prints a more accurate summary of the criterion used
for scf cycle termination.
Note that the criterion for total-energy convergence now uses the free
electronic energy, which is the appropriate magnitude to monitor in
the general case of possibly fractional occupancies.
The mixing flavor should be made into a first-class option variable, instead
of being encoded in two separate logicals (mix_charge, mixH)...
The new code uses a circular stack implemented using a "reference-counting"
paradigm.
|
Alberto Garcia |
10 years ago
|
|
|
| 454
|
|
|
Alberto Garcia |
10 years ago
|
|
|
| 453
|
|
|
Alberto Garcia |
10 years ago
|
|
|
| 452
|
|
|
Nick Papior Andersen |
10 years ago
|
|
|
| 451
|
|
|
Alberto Garcia |
10 years ago
|
|
|
| 450
|
|
|
Jose M Soler |
10 years ago
|
|
|
| 449
|
|
|
Nick Papior Andersen |
10 years ago
|
|
|
collapse all