[Wien] how to increase precision of calculations?

Martin Gmitra martin.gmitra at gmail.com
Thu Dec 11 15:11:18 CET 2014


Hi,

I had a look inside lapwso package and except a small redundancy in
rlomain.F, see lines 84 to 89,
<code snip>
                  bk(1)=bkrot(1)*br1(1,1)+bkrot(2)*br1(1,2)+bkrot(3)*br1(1,3)
                  bk(2)=bkrot(1)*br1(2,1)+bkrot(2)*br1(2,2)+bkrot(3)*br1(2,3)
                  bk(3)=bkrot(1)*br1(3,1)+bkrot(2)*br1(3,2)+bkrot(3)*br1(3,3)
                  BK(1)=BKROT(1)*BR1(1,1)+BKROT(2)*BR1(1,2)+BKROT(3)*BR1(1,3)
                  BK(2)=BKROT(1)*BR1(2,1)+BKROT(2)*BR1(2,2)+BKROT(3)*BR1(2,3)
                  BK(3)=BKROT(1)*BR1(3,1)+BKROT(2)*BR1(3,2)+BKROT(3)*BR1(3,3)
</code snip>
I have not found nothing that would solve my problem.

However, the non continuous behavior occurs once MBJ is turned on, see
attached plot of spin-orbit split valence band in wurtzite GaAs from
Gamma (k=zero) towards K point (shifted for a comparative purpose).
Just a note, that increasing IFFT do not heal the non-smooth behavior.

Any suggestions? Thanks in advance,
Martin


On Wed, Dec 3, 2014 at 12:04 PM, Martin Gmitra <martin.gmitra at gmail.com> wrote:
> Dear Peter,
>
> Thank you for your effort. I have modified lapwso and lapw1 so that
> 1/3 and 2/3 are entered in machine precision. The non-continuous
> behavior stays the same even with your suggestions for RKmax=10,
> Emax=9999.9 and adding RLOs in case.inso (could you comment please why
> consider RLOs in the GaAs case?). I had also changed the non-primitive
> translations exactly to 0 (did not mentioned last time) with no
> effect.
>
> I am observing that splitting goes perfectly smooth along the Gamma-H
> line. Therefore I plotted splitting along "Gamma-K" direction with a
> tiny kz component. Please see attached plots (idv=36000) for kz=0;
> kz=1/36000 and kz=5/36000. The smoothness gets slowly recovered. Could
> it be a helpful information to extract a reason for the problem?
>
> Best regards,
> Martin
>
> On Tue, Dec 2, 2014 at 12:03 PM, Peter Blaha
> <pblaha at theochem.tuwien.ac.at> wrote:
>> Hi,
>>
>> I could confirm the problem, but unfortunately I cannot offer a solution.
>>
>> These irregulations do not go away with increasing RKMAX or Emax (best is
>> anyway to set EMAX=9999.9 in case.in1 for such a study).
>> Changing compiler options (higher precision of ifort) or even changing to
>> gfortran or older WIEN2k versions does not cure the problem.
>>
>> They come from eigenvalue differences in the 9th digit after the comma.
>>
>> If I check internal accuracy, for instance by calculations for all 6
>> equivalent k-points (x,x,0;-x,-x,0;2x,-x,0,....) such differences should
>> still be feasible, because internal consistency is 11 digits in lapwso (but
>> 14 in lapw1, so clearly we are loosing something there), see below.
>>           58  0.449061656042513
>>           58  0.449061656068273
>>           58  0.449061656046884
>>           58  0.449061656046190
>>           58  0.449061656032196
>>           58  0.449061656067333
>>                           *
>>
>> In any case, for an accurate value of your quantity (delta-E/k) you need at
>> least
>>
>> RKMAX=10
>> Emax=9999.9  (all eigenvalues in lapw1)
>> add RLOs for both atoms in case.inso !!
>>
>> and then have to extrapolate from the smooth part of the curve.
>>
>> Change the non-primitive translation in case.struct to be exactly 0 !
>>
>> PS: another sophisticated test would be to convert your structure to an
>> orthorhombic one and test if the problem comes from the hexagonal lattice
>> (positions like 1/3, ... are never "correct" in the input), although then
>> you might have a problem to specify your k-points with full accuracy;  or at
>> least modify lapw1/lapwso and set the input 0.3333333 to 1/3 internally ..
>>
>>
>>
>> On 12/02/2014 08:22 AM, Martin Gmitra wrote:
>>>
>>> Dear Stefaan,
>>>
>>> Yes, I have increased the Emax parameter to 18 Ry, see blue curve in
>>> right plot. Indeed, it is one of the most important parameters.
>>>
>>> Best,
>>> Martin
>>>
>>>
>>> On Mon, Dec 1, 2014 at 1:20 PM, Stefaan Cottenier
>>> <Stefaan.Cottenier at ugent.be> wrote:
>>>>
>>>>
>>>>> Thanks for your reply and suggestions. I went with RKmax up to 12Ry
>>>>> and it turned out that 11 and 12 should be okay, since the splitting
>>>>> starts to "oscillate" (change in the increasing trend with increased
>>>>> RKmax). Concerning the non smooth behavior of the curve I did several
>>>>> test calculations playing around with the input parameters and none of
>>>>> the changes have brought desired healing of the step close to 0.003. I
>>>>> am attaching plots to the attachment.
>>>>
>>>>
>>>>
>>>>  From your mail, it's not clear whether or not you tried to increase
>>>> Emax, as
>>>> Fabien suggested. This is specified in the last line of case.in1, and is
>>>> different from RKmax which is at the top of that file. For spin-orbit
>>>> effects, this is the most important precision-determining feature.
>>>>
>>>> Stefaan
>>>>
>>>>
>>>>
>>>>> On Sun, Nov 30, 2014 at 12:02 AM,  <tran at theochem.tuwien.ac.at> wrote:
>>>>
>>>>
>>>>>>
>>>>>>
>>>>>> Actually, the most important is maybe the number of bands calculated
>>>>>> by lapw1 which are used for the 2nd variational procedure for the
>>>>>> spin-orbit coupling. So, to increase Emax in case.in1 (up to 20 Ry or
>>>>>> even
>>>>>> more) should be tried.
>>>>>>
>>>>>> F. Tran
>>>>>>
>>>>>>
>>>>>> On Sat, 29 Nov 2014, tran at theochem.tuwien.ac.at wrote:
>>>>>>
>>>>>>> Hi,
>>>>>>> For the convergence, I suggest to increase RKmax further. RKmax=8 is
>>>>>>> (very) good, but also not fully converged usually. Going up to
>>>>>>> RKmax=10 makes sense.
>>>>>>>
>>>>>>> It's difficult to say for the step at 0.003. Maybe the k-mesh is still
>>>>>>> not dense enough. An increase of GMAX (12 -> 18) in case.in2 is worth
>>>>>>> to
>>>>>>> try.
>>>>>>>
>>>>>>> F. Tran
>>>>>>>
>>>>>>> On Sat, 29 Nov 2014, Martin Gmitra wrote:
>>>>>>>
>>>>>>>> Dear Wien2k users,
>>>>>>>>
>>>>>>>> We are calculating spin-orbit coupling splitting of valence band in
>>>>>>>> wurtzite GaAs using mBJ. Since we are interested in a fine structure
>>>>>>>> close to Gamma point, we need to achieve sufficient accuracy. I am
>>>>>>>> attaching a plot of the valence band spin-orbit coupling splitting
>>>>>>>> (divided by the amplitude of k) as a function of k from the Gamma
>>>>>>>> point (k=0). As you can see, the splitting is sensitive to RKmax and
>>>>>>>> it does not look like it is going to "converge". What bothers me in
>>>>>>>> addition, is the step in the curves around 0.003 which can not be
>>>>>>>> healed by increasing RKmax (a smooth behavior as close as possible to
>>>>>>>> the Gamma point is desired result). I would like to add that
>>>>>>>> increasing k-mesh density does not affect the results as well as
>>>>>>>> increasing IFFT factor or reducing RMTs.
>>>>>>>>
>>>>>>>> Do you have any suggestions?
>>>>>>>> Best regards,
>>>>>>>> Martin Gmitra
>>>>>>>> Uni Regensburg
>>>>>>>>
>>>>>>>
>>>>>> _______________________________________________
>>>>>> Wien mailing list
>>>>>> Wien at 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
>>>>>>
>>>>>>
>>>>>> _______________________________________________
>>>>>> Wien mailing list
>>>>>> Wien at 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
>>>>
>>>>
>>>>
>>>> _______________________________________________
>>>> Wien mailing list
>>>> Wien at 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
>>>
>>> _______________________________________________
>>> Wien mailing list
>>> Wien at 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
>>>
>>
>> --
>>
>>                                       P.Blaha
>> --------------------------------------------------------------------------
>> Peter BLAHA, Inst.f. Materials Chemistry, TU Vienna, A-1060 Vienna
>> Phone: +43-1-58801-165300             FAX: +43-1-58801-165982
>> Email: blaha at theochem.tuwien.ac.at    WIEN2k: http://www.wien2k.at
>> WWW:   http://www.imc.tuwien.ac.at/staff/tc_group_e.php
>> --------------------------------------------------------------------------
>> _______________________________________________
>> Wien mailing list
>> Wien at 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
-------------- next part --------------
A non-text attachment was scrubbed...
Name: PBE_vs_MBJ_SOC_split.png
Type: image/png
Size: 12157 bytes
Desc: not available
URL: <http://zeus.theochem.tuwien.ac.at/pipermail/wien/attachments/20141211/64527e27/attachment-0001.png>


More information about the Wien mailing list