VASP errors on Mendel

[Saurabh Bajaj]

Hi all,

I was running defect calculations using VASP on Mendel, and for some reason, the structure optimization process results in the following warning, repeated several times, and the relaxation process getting stuck in a particular ionic step and oscillating back and forth between large deviations in energy. 

 WARNING: Sub-Space-Matrix is not hermitian in DAV            78

   -15.42803679090842

Below I have pasted the INCAR file I am using. I ran the same calculation on a different cluster with the same version of VASP (5.2.2) and it gives me no issues there. So I am not sure if this error has something to do with the libraries being used on Mendel, or something else. A little Googling didn't result in any concrete answers. Anyone have any ideas regarding this?

Thanks in advance

Saurabh

=====================INCAR=======================

SYSTEM = UO2 bulk

 Startparameter for this Run:

  ISTART =  0;       job   : 0-new  1-cont  2-samecut

 Electronic Relaxation 1

  ENCUT  =  400.0      eV(25.7 Ryd)

  NELM   =  500        number of electronic steps between two ionic steps

  EDIFF  =  1E-05      stopping-criterion for ELM

  PREC = Normal

  LWAVE = .FAlSE.

  LCHARG = .FALSE.

  SYMPREC = 1E-3

 DOS related values:

  ISMEAR = 1

  SIGMA = 0.1

  ICHARG = 2

 # ISPIN = 2

 # MAGMOM = 8*5

 # LMAXMIX = 6          ! Important: mix paw occupancies up to L=4

 # LORBIT = 11

  IDIPOL = 4

  EPSILON = 593.67

 Electronic Relaxation 2

 # ALGO = Fast

  LREAL  =   Auto     real-space projection

 # NBANDS  =  80

  NELECT = 2492

 Ionic Relaxation

#  EDIFFG = 1E-05

  NSW    =      100   number of steps for IOM

  NBLOCK =       1    KBLOCK =      1    inner block; outer block

  IBRION =       2    ionic relax: 0-MD 1-quasi-New 2-CG

  ISIF   =       3    stress and relaxation

 # IWAVPR =      1    prediction:  0-non 1-charg 2-wave 3-comb

 # LCORR  =      T    Harris-correction to forces

[Anubhav Jain]

Hi Saurabh

Generally we set LREAL=False if we get this error; if you run through custodian this should typically happen automatically

--
You received this message because you are subscribed to the Google Groups "Materials Project Development Group" group.
To unsubscribe from this group and stop receiving emails from it, send an email to matproj-devel...@googlegroups.com.
To post to this group, send email to matproj...@googlegroups.com.
Visit this group at http://groups.google.com/group/matproj-develop.

[Saurabh Bajaj]

Hi Anubhav,

Thanks for your reply. I set LREAL = False and re-ran the calculation but it seems to have slowed down the calculation tremendously - in almost 24 hours of run time on 1 node (16 cores), it hasn't even converged to the first ionic step. Would you know of any other reason that might be causing these errors?

Thanks

Saurabh

[Saurabh Bajaj]

Just a quick update - using 'vasp/5.3.3_vtst.matgen' with 'LREAL  = Auto' resulted in normal behavior and avoided running into any errors.

[Anubhav Jain]

Interesting...I wonder what changed...

I would still double-check the results are actually converged at the end, i.e. by inspecting the OSZICAR and looking at the number of iterations

[Saurabh Bajaj]

Yeah, I've been looking at the iterations, and so far it looks good.

[Xi Jianqi]

Hi, I meet the same issue. The LREAL=FALSE seems will take a long time for my supercell calculation. Where can I get the vasp/5.3.3_vtst.matgen version? Thanks

在 2015年4月28日星期二 UTC-4上午10:07:27，Saurabh Bajaj写道：

[Saurabh Bajaj]

It doesn't seem to be available on matgen now. Edison has many more versions, including the "vtst" custom builds of vasp. Maybe you could try one of those. 

[Xi Jianqi]

Thanks very much! I will try vtst.

在 2016年7月17日星期日 UTC-4下午2:25:37，Saurabh Bajaj写道：