[Wien] RMT, El and warnings in qtl-b and :ENE

[Stefaan Cottenier]

See the summary at the end as well.

>I would like to obtain a good set of linearization energies (case.in1)
>for fcc nickel and other geometries, where I closely follow the paper
>by Dr. Cottenier.
>
>  
>
Just a warning: in many cases the default values are fine, and the 
'fine-tuning' by -in1new or by inspection of the DOS does not add 
meaningful improvement. Only if you have slabs and/or heavy atoms (which 
means: cases where :FER is far away from a typical value as 0.5 Ry), 
this becomes important.

>Here is what I did. All the calculations are run sequentially on a
>single computer (intel, see below) and parallel calculations on a
>different machine (Opteron) do not have the following problem.
>  
>
I assume you have exactly the same wien2k version on both machines?

>(1) Reproduce the fccni results (example), with default parameters in
>case.in1 and sp case plus -cc 0.0001 (I run first -ec 0.00001).
>Results are good.
>
>userguide: MMTOT: 0.63068 (RMT=2.3 )
>mine            : 0.63041 (with -ec RMT=2.35)
>                  0.63040 (with -cc RMT=2.35)
>                  0.63066 (with -cc RMT=2.30)
>
>  
>
Perfect -- this is numerical noise only.

>(2) Change rkm from 7.0 to 9.5 in case.in1 (3000 KPT) by
>re-initializing (init) (case.in1 is checked afterwards to make sure
>9.5 is there and save all the  previous calculation (save))
>
>  (a) If RMT=2.30, then
>
>         my         MMTOT:   0.62917  (runsp -I -i 40 -ec 0.00001  )
>          optimized MMTOT:   0.63057  (runsp -I -i 40 -ec 0.00001 )
>            ^__no qtl-b warning and no :ENE **WARN** for energy
>  (b) If RMT=2.35 (which is suggested by the program), then
>	FORTRAN STOP  LAPW0 END
>	FORTRAN STOP SECLR4 - Error
>         Cholesky INFO =           91
>         'SECLR4' - POTRF (Scalapack/LAPACK) failed.
>
>(3) Does this (referring to 2(b)) indicate overcompleteness as
>suggested by Dr. Cottenier, where the quasi-linear dependence among
>the basis states as you mentioned? If so, why does my parallel job on
>a pc cluster run without any problem?  Is it possible each machine has
>different thresholds to invoke the Cholesky warning?
>  
>
9.5 is a very large basis set, while 3000 k-points (I assume in the full 
BZ?) is a rather coarse mesh for a metal. To my experience, this SECLR4 
error often (but not always) disappears if the k-mesh is improved. If 
there is no good reason to go to 9.5, stay away from it. For 'routine 
work' you rarely need more than 7.5 with APW+lo.

>(4) I noticed in attached fccni example, if  grep :ENE *.scf
>
>:ENE  : ********** TOTAL ENERGY IN Ry =        -3041.652575
>:ENE  : ********** TOTAL ENERGY IN Ry =        -3041.652605
>
>however, if I use the exact same parameters (case.in1,in2,struct), I got
>
>:ENE  : *WARNING** TOTAL ENERGY IN Ry =        -3041.652612
>:ENE  : *WARNING** TOTAL ENERGY IN Ry =        -3041.652589
>
>even though I got a similar qtl-b 3.72 or 2.47.
>
>Is this normal or something I should worry?
>
>  
>
That's why I ask whether you are running exactly the same wien2k 
versions on both systems. The treshold to give a qtl-b warning has been 
changed a few times (if I remember well).

>(5) I managed to get rid of the above warnings (qtl-b and *WARNING*)
>by modifying the case.in1's parameters (E_l), but I found it is not
>enough to simply follow Dr. Cottenier's method which emphasizes on the
>low-lying states (below Ef) and I have to combine -in1new and dos to
>get rid of warnings of states above Ef. I wish someone would
>systematically explain how to balance the unoccupied and occupied
>states, and in particular how to add additional E_l such as 2 1.50
>0.000 Cont 0. In Ni, adding this line leads to an immediate error and
>make things worse. Is there any rule of thumb or theory behind this?
>  
>
I just verified that adding a line as you described runs fine for Ni 
(perhaps you did not change the 'number of lines' from 4 to 5?)

To summarize: I do not think you really have a problem. Ni is a bit 
'special' as it has quite a large exchange splitting of up and down 
bands. This makes it difficult to find 1 d-linearization energy for both 
bands (in1new suggests 0.4, with the peaks of the bands being at 0.2 and 
0.6, roughly). That leads to completely innoncent qtl-b warnings (as 
discussed before in the mailing list, APW+lo can tollerate such small 
qtl-b's).

Stefaan


Disclaimer: http://www.kuleuven.be/cwis/email_disclaimer.htm