[Wien] proper k-points for Nd
Peter Blaha
pblaha at theochem.tuwien.ac.at
Tue Jan 22 22:00:45 CET 2019
Please, do NOT use iterative diagonalization !!!!
Iterative diagonalization is for "few eigenvalues and large cells" (like
surfaces, or supercells where every atoms has only few electrons.
Iterative diagonalization is NOT for f-systems !!
Firstly, you will not get any speedup, but even worse, the
diagonalization and scf-convergence may not finish at the same time,
i.e. your eigenvalues/verctors are wrong.
(-it explains why you get ghostbands when changing the k-mesh (should
use -it1 or touch .fulldiag), and also explains why you get so different
energies. These numbers are simply not converged.
Am 22.01.2019 um 20:34 schrieb Victor Zenou:
> Thanks Peter and Laurie for your advises. I hope we all stay friends
> after my finding :)
> The reason I didn't progress with higher k-points (I guess it doesn't
> work backwards, higher k-points to lower) is that I actually tried that
> few times and it worked (at other phase),but sometimes I got error
> message after the first cycle. In order to be systematic I tried to
> avoid that.
> According to Peter's advice I tried again and got an error message (I
> can't recall if it was the same as previous) at lapw2 -up of the first
> cycle. uplapw2.error file showed: " 'LAPW2" - semicore band-ranges too
> large, ghostbands ? ". I 've deleted that file and re- run SCF while
> unmarking the "iterative diag" box and for some reason it worked. It
> worked another time, as I run 2 parallel sessions/cases with different
> k-points (Nd-SP and Nd-SP2).
> I have good news and bad ones. The good news
> 20,000 k-points gave the same energy as 15,000 (ec 0.0001, cc 0.0001):
> 77040.469768(3) and 77040.469768(2), with same MMTOT (16.645, 16.647).
> :CHARGE convergence -.000216 and -.000422, for 15000 and 20000 k-points!
>
> Now for the troubling news:
> 30,000 k-points gave something different -77040.506092(3), MMTOT=16.091;
> CHARGE convergence -.000070.
> Running the same at different case, as I said I run 2 parallel cases
> using the same case.struc file) I got for the same 30,000 k-points
> -77040.4874669(3), MMTOT=16.682; CHARGE convergence -.000070.
> It seems that using same k-points gave me two different energies. As I
> understand repeating calculation should give the same results. And if
> these two calculations are the same, unless I did something stuoid,
> should give the same result. Also, the different MMTOT may suggest
> something wrong with the spin polarized calculations. Maybe I should do
> the k-test without spin polarization, as I usually do.
> I used both to try 60,000 points, and later I will follow Laurie's and
> Peter HDLOs.
> Victor
>
>
>
> בתאריך יום א׳, 20 בינו׳ 2019 ב-20:01 מאת Peter Blaha
> <pblaha at theochem.tuwien.ac.at <mailto:pblaha at theochem.tuwien.ac.at>>:
>
> I disagree with Lauries point: Eventually I want to do even an
> "unphysical" calculation, of course it might be you have to pay a price
> for it, because this unphysical model as eg. a spike in the DOS at EF.
> And mixer should lead to convergence, but sometimes has problems
> (Lauries mixer is still one of the big advantages of WIEN2k. I know
> other codes, which would not be able to converge a 5f system at all).
>
> In any case, put a HDLO (5f) and increase LVNS=6 first.
>
> Most importantly: why are you starting from scratch when doing the
> k-mesh tests. This is very error prone and in particular burns
> cpu-hours
> completely unnecessarily.
>
> One of my main "advises" is always: do an init_lapw only ONCE ! (except
> you have to change spheres).
>
> You will get much better internal convergence when you continue using
> the previous calculation:
>
> loop:
> runsp ....
> save XXX-k-points
> x kgen (increase the mesh always at least by a factor of 2 (maybe
> even 4).
> goto loop
>
> I'm not so much surprised about the k-mesh for a magnetic 5d-element
> calculation. I start with 10000 k-points for every metallic spd-element
> and test it with 50000 k. You may need 100 000 or even more. (optics
> for
> fcc Al is not fully converged with 1 000 000 k-points !). It does not
> cost much cpu, since you should start with a nearly converged
> calculation.
>
> Your :DIS is still quite large, try to bring it below 0.0001
>
> You can try TEMPS 0.002 as Laurence advised. This will most likely
> improve convergence, but most likely also reduce MMT a little bit,
> which
> is ok when you compare with room temp experiments, but not for 4k
> measurements ...
>
>
>
> Am 20.01.2019 um 17:50 schrieb Laurence Marks:
> > Sorry Fabien, I am dubious that k-points is the issue, I suspect
> a bad
> > model.
> >
> > Let me explain, repeating some things I have said before or are
> in the
> > mixer notes albeit perhaps buried.
> >
> > An "easy" assumption that many people make is that dft problems
> are well
> > posed, so have "simple" fixed-point solutions. In fact this is
> not true.
> > In reality dft problems have stacks of numerical issues which
> leads to
> > "noise". You can think about this in terms of finding the minimum of
> > "F(x) = x*x + Constant*Noise(x)" where Noise(x) is some randomly
> > generated noise. The larger Constant is, the larger the deviation is
> > from x=0 of the apparent minimum. Averaging over more k-points may
> > reduce Constant, but often in my experience does relatively little.
> >
> > A symptom of noise is poor convergence -- which is clearly what is
> > occurring here.
> >
> > If you improve the convergence rate you will reduce the noise. Some
> > suggestions:
> >
> > 1) Use HDLO and reduce RMT
> > 2) Reduce RMT and use HDLO
> > 3) Larger LM (e.g. 8)
> > 4) Larger gmax in case.in2 (e.g. 16)
> > 5) Use TEMPS 0.0018 (room temperature)
> > 6) Peter's suggestions
> >
> > On Sun, Jan 20, 2019 at 10:27 AM tran at theochem.tuwien.ac.at
> <mailto:tran at theochem.tuwien.ac.at>
> > <mailto:tran at theochem.tuwien.ac.at
> <mailto:tran at theochem.tuwien.ac.at>> <tran at theochem.tuwien.ac.at
> <mailto:tran at theochem.tuwien.ac.at>
> > <mailto:tran at theochem.tuwien.ac.at
> <mailto:tran at theochem.tuwien.ac.at>>> wrote:
> >
> > Hi,
> >
> > If not too expensive, you could try more k-points (20000 and
> 25000)
> > to see if the fluctuations persist.
> >
> > Questions: how many atoms in the unit cell? Are you
> optimizing the
> > position of atoms?
> >
> > F. Tran
> >
> > On Sunday 2019-01-20 17:17, Victor Zenou wrote:
> >
> > >Date: Sun, 20 Jan 2019 17:17:55
> > >From: Victor Zenou <zanov at post.bgu.ac.il
> <mailto:zanov at post.bgu.ac.il>
> > <mailto:zanov at post.bgu.ac.il <mailto:zanov at post.bgu.ac.il>>>
> > >Reply-To: A Mailing list for WIEN2k users
> > <wien at zeus.theochem.tuwien.ac.at
> <mailto:wien at zeus.theochem.tuwien.ac.at>
> > <mailto:wien at zeus.theochem.tuwien.ac.at
> <mailto:wien at zeus.theochem.tuwien.ac.at>>>
> > >To: A Mailing list for WIEN2k users
> > <wien at zeus.theochem.tuwien.ac.at
> <mailto:wien at zeus.theochem.tuwien.ac.at>
> > <mailto:wien at zeus.theochem.tuwien.ac.at
> <mailto:wien at zeus.theochem.tuwien.ac.at>>>
> > >Subject: Re: [Wien] proper k-points for Nd
> > >
> > >Dear Peter, LDM and wien2k users
> > >I did some more calculation with energy and CHARGE convergence
> > criteria
> > >equal to 10^-4 Ryd and 10^-3, respectively. The energy usually
> > converged to
> > >1*10-5 or even 1*10^-6. Still I get energy fluctuation up to
> > 4*10-2 Ryd
> > >(please see below)
> > >I didn't check AFM yet, but this is not the point. The
> point is to get
> > >"stable" calculation for optimal k-points.
> > >I have to say that for each calculation I totally aresed
> all file
> > except
> > >case.struc.
> > >Also I used all defaults as well as gmax=12, also
> > rmt(min)*kmax=8.5, for all
> > >calculations.
> > >I can repeat all calculation with smaller RMT, such as 2.5,
> or to
> > check
> > >HDLO.
> > >I think I will start with smaller RMT. Any comments /
> suggestions?
> > >Victor
> > >
> > >K-pointsIBZ E [Ryd], 10-4 CHARGE convergence CHARGE
> DISTANCE MMTOT
> > >3000 208 -77040.50516761 -.000504 0.000825 16.00318
> > >5000 280 -77040.48748499 -.000544 0.000758 16.68523
> > >7500 455 -77040.51668342 -.000862 0.000183 16.62547
> > >10000 540 -77040.48747210 -.000817 0.000499 16.67544
> > >15000 840 -77040.46976815 -.000216 0.001362 16.64536
> > >
> > >
> > >
> >
> >
> >
> > --
> > Professor Laurence Marks
> > "Research is to see what everybody else has seen, and to think what
> > nobody else has thought", Albert Szent-Gyorgi
> > www.numis.northwestern.edu <http://www.numis.northwestern.edu>
> <http://www.numis.northwestern.edu> ;
> > Corrosion in 4D: MURI4D.numis.northwestern.edu
> <http://MURI4D.numis.northwestern.edu>
> > <http://MURI4D.numis.northwestern.edu>
> > Partner of the CFW 100% program for gender equity,
> > www.cfw.org/100-percent <http://www.cfw.org/100-percent>
> <http://www.cfw.org/100-percent>
> > Co-Editor, Acta Cryst A
> >
> > _______________________________________________
> > Wien mailing list
> > Wien at zeus.theochem.tuwien.ac.at
> <mailto: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
> >
>
> --
> --------------------------------------------------------------------------
> 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
> <mailto:blaha at theochem.tuwien.ac.at> WIEN2k: http://www.wien2k.at
> WWW:
> http://www.imc.tuwien.ac.at/tc_blaha-------------------------------------------------------------------------
>
>
> _______________________________________________
> Wien mailing list
> Wien at zeus.theochem.tuwien.ac.at <mailto: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
>
--
--------------------------------------------------------------------------
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/tc_blaha-------------------------------------------------------------------------
More information about the Wien
mailing list