Dear Laurence,

thank you for your detailed answer.

I have tried all your suggestions,
- I changed case.in0 with increased oversampling by factor two (new parameters LUSE 26 and IFFTfactor 4)
------------ start of case.in0 -----------
TOT  XC_LDA (XC_PBE,XC_PBESOL,XC_WC,XC_MBJ,XC_REVTPSSS)
NR2V      IFFT  26  (R2V)
   24   24   24    4.00  1    min IFFT-parameters, enhancement factor, iprint
------------ end of case.in0 -----------

- I also tried to impose strong convergence criteria to be -cc 0.00000001 -ec 0.00000001

However, in both cases, the ghost MLD remains practically identical as when using my default values (default Fe + convergence -cc 0.00001 -ec 0.000001)

Also, final :PUP 'Current' parameters remained practically the same for all the calculations (by about 2 digits), being like (for M001)
PW CHANGE     H    K    L      Current       Change    Residue
:PUP001:      0    0    0  2.10719649E-02  5.090E-10 -7.530E-09
:PUP002:      0   -1   -1  3.67084665E-04 -7.004E-10 -3.572E-10
:PUP003:      1   -1    0  1.82124988E-04 -3.669E-10 -2.211E-10
:PUP004:      0    0   -2 -1.87471938E-03  6.192E-11  7.254E-10
:PUP005:      0   -2    0 -3.75090680E-03  1.125E-10  1.378E-09
:PUP006:      1   -1   -2 -3.46372731E-03  1.026E-10  1.378E-09
:PUP007:      1   -2   -1 -6.92804324E-03  2.418E-10  2.776E-09
:PUP008:      0   -2   -2 -7.14014083E-04  8.677E-11  2.032E-10
:PUP009:      2   -2    0 -3.57036516E-04  4.298E-11  1.121E-10
:PUP010:      0   -1   -3  2.62159615E-04  9.199E-11 -1.588E-10
:PUP011:      0   -3   -1  2.62289930E-04  9.844E-11 -1.555E-10
:PUP012:      1   -3    0  2.62219244E-04  9.133E-11 -1.551E-10
Unfortunately, all reflections seems to be allowed for bcc (H+K+L is even), forbidden reflections of bcc are (H+K+L=odd), so I can not see how they get close to zero ;-)) But the idea is excellent.

Thank you again and with my best regards

Jaroslav

On 27/11/17 15:27, Laurence Marks wrote:
Let me clarify slightly my comment about symmetry -- as I realized the explanation (I think) and can also suggest something that might help.

First, concerning symmetry the explanation is I believe simple. If the problem has a real symmetry operation such as inversion which is being removed, then the Jacobian at the solution has zero's for charge disturbances that break this symmetry. Because of this noise due to numerical accuracy has a large effect, and almost certainly one has to tighten the convergence criteria particularly -cc. You can monitor this by looking at the :PUPXXX values in case.scfm and look how well the forbidden reflections have converged to zero.

Second, do not be surprised about numerical issues. While the calculations are done in double precision, there are many large sums and in some cases double sums, and also numerical integrations/differentiation. Any large sum or numerical integration/differentiation in general reduces the numerical accuracy. Hence even though double precision has an accuracy of 1D-15 the sum may only be accurate to 1D-10 or even 1D-7. Also, the Intel ifort compiler will reduce the numerical accuracy for speed if one is not careful.

One thing that may help is to increase the oversampling in case.in0 for VXC, both that of the PW's and of the CLMs. A standard test is to use LDA and see if the problem goes away, since oversampling is much less relevant for this.

Of course your problem may have nothing to do with any of this....

_______________________________________________
Wien mailing list
Wien@zeus.theochem.tuwien.ac.at
http://zeus.theochem.tuwien.ac.at/mailman/listinfo/wien
SEARCH the MAILING-LIST at:  
http://www.mail-archive.com/wien@zeus.theochem.tuwien.ac.at/index.html

Reply via email to