Page 1 of 2

SCF convergence for a big system

Posted: Wed Feb 29, 2012 12:39 pm
by quan.phung
Dear Vasp user,

I am a newbie with VASP. I am trying to do a calculation with a big system, contains 160 atom. I have a problem that the SCF never converges and dies after several cycles:

energy without entropy =****************** energy(sigma->0) =******************

This is the INCAR file
general:
System = YGG
PREC = Med
ISTART = 0 ; ICHARG=2
ENCUT = 520
ISMEAR = 0; SIGMA = 0.1;
LREAL = Auto ! Projection done in real space
ALGO = Fast
AMIX = 0.2
NELM = 100

Do you have any suggestion for my input file to solve the problem?

Best regards,
Quan

SCF convergence for a big system

Posted: Wed Feb 29, 2012 5:41 pm
by Danny
are you sure it dies, and not just is bussy doing calculations...just very slowly due to the size of the system (maybe even in combination with massive swapping) --> wrt this, are you sure your calculation "can" run on the machine you run it on (enough memory etc)

SCF convergence for a big system

Posted: Wed Feb 29, 2012 9:21 pm
by quan.phung
Dear Danny,

I'm sure that the calculation dies with something like:
Error EDDDAV: Call to ZHEGV failed.
I "partially" solved the problem by changing SIGMA = 0.01 even my system is a semiconductor. But the SCF now is very slow.
I used "partially" because when I try to do kpoint convergence study, with 1x1x1 or 3x3x3 mesh, the calculation converged, but it died with 2x2x2 and 4x4x4 mesh.
For the machine, I used 26 cores with 20 GB Ram and I had no problem with memory.

Best regards,
Quan

Re: SCF convergence for a big system

Posted: Fri Sep 11, 2020 4:30 am
by SKM
hi
though its a very past isntance of issue, may i know if the issue is solved?
i have a 224 atom supercell of perovskite and not electronic converging is not happening. Tried NELM increase just to see if it anyway converges. No use. Meddled with mixing parameters, No success. Reduced POTIM as in some forum suggested, No success. i have to do IFIF=3, as i want to check the cell lattice changes of doping. not using any MAGMOM tag. Does it effect the SCF convergence? any clue? Eagarly waiting for a response from experienced, please.

Re: SCF convergence for a big system

Posted: Fri Sep 11, 2020 11:41 am
by merzuk.kaltak
Please upload your INCAR, POSCAR, POTCAR, KPOINTS and OUTCAR file inkluding the std in form of a zip file.

Re: SCF convergence for a big system

Posted: Wed Sep 23, 2020 6:49 am
by SKM
Hi
Thanks for the reply. subsequent to my query i did many trails, with different options etc, and discussed in Researchgate forum, and if you kindly see the URL, (satrting from sept 2020) you might get idea what all tried, and may help you to resolve the issue.
https://www.researchgate.net/post/How_c ... 750c13ad57

Regarding the files you asked, can they be protected? as some input is from a collaborator.
Regards

Re: SCF convergence for a big system

Posted: Wed Sep 23, 2020 7:43 am
by merzuk.kaltak
No, they are not protected. All uploaded files on this forum are accessible by registered users.
That is, all vasp users with a valid license can access data uploaded here.

Re: SCF convergence for a big system

Posted: Thu Sep 24, 2020 2:03 am
by SKM
Thanks for the reply. attached the .tar.gz file. Please note the INCAR file you see is the latest used one with those mixing parameters commented out. But otherwise i tried with those values that were commented out and also different values for AMIX and BMIX in different runs/trails.
i tried increasing NELM to even 640. tried IBRION=2, and IALGO=38 or 48 etc.
all such trials which i can think of and based on some web forum suggestions etc. were done.
Hope you help to resolve this issue. as once its resolved i will have another similar system with > 600 atoms.

Regards

Re: SCF convergence for a big system

Posted: Wed Sep 30, 2020 12:32 pm
by henrique_miranda
You should start by simplifying the problem and step by step increase the complexity until you have the problem you want to solve.

In particular:
- start by doing a gamma only calculation with vasp_gam (it will run much faster so you will know much faster if something is wrong).
- Your electronic calculation is not converging so there is not point in doing a ionic relaxation as the forces will most likely be wrong. Start by getting a converged electronic calculation, only them its reasonable to add the relaxation variables (IBRION,ISIF,NSW). Normally the electronic part converges in a few iterations ~30 if it does not converge in 30 iterations you should at least look at the total energy at each step to see what is happening (in your case it is oscillating a lot).
- Try to comment some of the input variables (ISYM, ENCUT, NELMDL, NELM, NELMIN, PREC) to figure out if it solves your problem. VASP default values for the inputs work for many systems, you should only specify these parameters once you are sure of what they do. If the calculation works with these variables commented, then you can uncomment them one at a time until you find which one was causing your problem.
- Your system is metallic, those can be a bit more difficult to converge. You should make sure that you have enough empty states in your calculation: reduce SIGMA, increase NBANDS look into the ISMEAR options (1 and -1 should be good choices for metals). Looking to the OUTCAR it seems you have enough empty bands though...

PS: you did not post the POSCAR so its hard to reproduce the issue you are having.
I removed the files you posted previously because a POTCAR file was present and it should not according to the forum rules (leave only the reference to the POTCARS you used, i.e. the first line of each POTCAR).

There is not a single reason why a calculation does not converge and naturally not a single solution.
You have to try different things until you understand why it does not work.

Re: SCF convergence for a big system

Posted: Wed Sep 30, 2020 2:09 pm
by SKM
Hi
Thanks for the reply. Oh sorry. How could I forget the POSCAR.? by mistake !!! Do you need it now? or first i will check the points one by one what you suggested and get back.
and by the way,

How could we understand the sytem is metallic or not or something?

Regards.

Re: SCF convergence for a big system

Posted: Wed Sep 30, 2020 9:21 pm
by henrique_miranda
The system is metallic as long as there are states that are partially occupied.
You can check that in the OUTCAR file.

Re: SCF convergence for a big system

Posted: Fri Oct 02, 2020 8:17 am
by SKM
Hi
incidentally, before your previous suggestion/s, i did a Gamma only run (Run-1), but not converged.
After your suggestion also, i did Gamma only run (Run-2) by commenting many tags in INCAR. But in Run-2, the OSZICAR file gives the summary output line similar to AIMD run, surprisingly, but i dont think iused any MD tags.

Kindly suggest what is going on in the runs, (attached two tar files for Run 1 & 2).

Regards

Re: SCF convergence for a big system

Posted: Fri Oct 02, 2020 9:22 am
by henrique_miranda
I suggested:
Start by getting a converged electronic calculation, only them its reasonable to add the relaxation variables (IBRION,ISIF,NSW).
Look at your OUTCAR-1 file: IBRION is set to 2 which means ionic relaxation (even though the input file has IBRION commented so you are probably using a different input file than the one you sent)
Look at your OUTCAR-2 file: IBRION is set to 0 which means MD run

Simplify the problem: try to obtain a converged electronic structure calculation.
Another suggestion: look at your structure (POSCAR) with VESTA or some similiar software, see if everything is like what you expect.
- Your system is metallic, those can be a bit more difficult to converge. You should make sure that you have enough empty states in your calculation: reduce SIGMA, increase NBANDS look into the ISMEAR options (1 and -1 should be good choices for metals). Looking to the OUTCAR it seems you have enough empty bands though...
Did you try both these options for ISMEAR?
Did you try different values for SIGMA?

As I said before: convergence issues are complicated. Not a single recipe works everywhere.
Lots of testing is needed, different incar tags and combinations have to be tested but you have to be the one doing the testing.
Please post here results from your tests, I or someone else will help you track down the issue :)

Re: SCF convergence for a big system

Posted: Sun Oct 25, 2020 1:58 pm
by SKM
Hi
Sorry for late comeback on the issue and messing the INCAR files submitted. Actually, what I attached were those edited after the run, so messed-up. Kindly ignore.
I am trying to understand what you mentioned
Start by getting a converged electronic calculation, only then its reasonable to add the relaxation variables (IBRION,ISIF,NSW)
i did two tests with default NELM and with 200 NELM. kindly see the attachments of the runs. i think i still did not get what you mentioned above and how to do it.

Re: SCF convergence for a big system

Posted: Sun Nov 08, 2020 11:04 am
by SKM
Dear Admin,

kindly look into this and help.
Regards