Problems with structural relaxation with VASP 6.5.* and IBRION=2

[dholec]

Dear VASP developers,

Recently I came accross a strange behaviour during structural relaxation using VASP 6.5.0 with IBRION=2: the system seemed to be almost converged (total energy changes ~-10e-3) when all of a sudden positive (!) energy change >~+10e+0 was reported in OSZICAR for about 2-3 ionic steps, after which VASP finishes without any error message. Visualizations of the structural changes (XDATCAR) do not show any visible changes in the atomic positions in the last ionic steps.

This behaviour was also reproduced with VASP 6.5.1. When switching to older versions (VASP 6.4.2 and VASP 6.4.3), the convergence worked as expected. Strangely enough, I also get the correct results with the new VASP versions when using IBRION=1.

I will appreciate any insight into what may be going on. Thank you very much in advance.

David Holec

---

In short:
vasp 6.4.2, vasp 6.4.3 - converge as expected
vasp 6.5.0, vasp 6.5.1 - converge only with IBRION=1

---

Report from the calculations extracted with:

Code: Select all

for v in v6*; do 
  echo $v
  head -n 1 $v/OUTCAR
  grep "IBRION" $v/OUTCAR | tail -n 1
  grep IALGO $v/OUTCAR
  grep "F" ${v}/OSZICAR | tail -n 3
  grep accuracy $v/OUTCAR
  echo ""
  echo "------"
done

(module = binary provided by the cluster admins
own, old, new = various self compilations (gcc) using own compiled Scalapack, BLAS, FFTW libraries as well as those taken from modules available on the cluster)

Code: Select all

v642_module
 vasp.6.4.2 20Jul23 (build Feb  8 2024 14:03:34) complex                         

   IBRION =      2    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
  26 F= -.57739808E+03 E0= -.57739830E+03  d E =-.446683E-03  mag=   164.0169
  27 F= -.57739812E+03 E0= -.57739835E+03  d E =-.376100E-04  mag=   164.0136
  28 F= -.57739814E+03 E0= -.57739835E+03  d E =-.589476E-04  mag=   164.0103
 reached required accuracy - stopping structural energy minimisation

------
v642_module_IBRION1
 vasp.6.4.2 20Jul23 (build Feb  8 2024 14:03:34) complex                         

   IBRION =      1    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
   9 F= -.57739597E+03 E0= -.57739605E+03  d E =-.260357E-03  mag=   164.2198
  10 F= -.57739613E+03 E0= -.57739626E+03  d E =-.165832E-03  mag=   164.2138
  11 F= -.57739619E+03 E0= -.57739631E+03  d E =-.570777E-04  mag=   164.2045
 reached required accuracy - stopping structural energy minimisation

------
v642_own
 vasp.6.4.2 20Jul23 (build Feb  6 2024 11:48:26) complex                         

   IBRION =      2    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
  20 F= -.57739724E+03 E0= -.57739747E+03  d E =-.159384E-03  mag=   164.1169
  21 F= -.57739755E+03 E0= -.57739777E+03  d E =-.309319E-03  mag=   164.1133
  22 F= -.57739741E+03 E0= -.57739763E+03  d E =-.162993E-03  mag=   164.1081
 reached required accuracy - stopping structural energy minimisation

------
v643_own
 vasp.6.4.3 19Mar24 (build Apr  3 2024 11:27:27) complex                         

   IBRION =      2    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
  16 F= -.57739701E+03 E0= -.57739715E+03  d E =-.175294E-03  mag=   164.1739
  17 F= -.57739709E+03 E0= -.57739725E+03  d E =-.784421E-04  mag=   164.1660
  18 F= -.57739710E+03 E0= -.57739729E+03  d E =-.886936E-04  mag=   164.1542
 reached required accuracy - stopping structural energy minimisation

------
v650_new
 vasp.6.5.0 16Dec24 (build Aug  8 2025 16:25:17) complex                         

   IBRION =      2    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
  12 F= -.10683552E+04 E0= -.10683576E+04  d E =-.490960E+03  mag=   166.9632
  13 F= -.16489782E+04 E0= -.16489768E+04  d E =-.107158E+04  mag=   146.3111
  14 F= -.29843258E+04 E0= -.29843224E+04  d E =-.240693E+04  mag=    86.0477

------
v650_orig
 vasp.6.5.0 16Dec24 (build Jan 23 2025 14:48:21) complex                         

   IBRION =      2    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
   8 F= -.57739482E+03 E0= -.57739488E+03  d E =-.588512E-03  mag=   164.2412
   9 F= -.55973092E+03 E0= -.55973089E+03  d E =0.176633E+02  mag=   164.1832
  10 F= -.57704163E+03 E0= -.57704170E+03  d E =0.352602E+00  mag=   164.1977
 reached required accuracy - stopping structural energy minimisation

------
v650_orig_IBRION1
 vasp.6.5.0 16Dec24 (build Jan 23 2025 14:48:21) complex                         

   IBRION =      1    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
   6 F= -.57739511E+03 E0= -.57739516E+03  d E =-.566428E-03  mag=   164.2557
   7 F= -.57739561E+03 E0= -.57739565E+03  d E =-.497384E-03  mag=   164.2464
   8 F= -.57739571E+03 E0= -.57739578E+03  d E =-.995651E-04  mag=   164.2315
 reached required accuracy - stopping structural energy minimisation

------
v651
 vasp.6.5.1 10Mar25 (build Aug  9 2025 08:30:21) complex                         

   IBRION =      2    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
  10 F= -.48936945E+03 E0= -.48937140E+03  d E =0.880238E+02  mag=   124.4836
  11 F= -.65829768E+03 E0= -.65829839E+03  d E =-.809044E+02  mag=   105.7547
  12 F= -.19291716E+04 E0= -.19291718E+04  d E =-.135178E+04  mag=    78.1989

------
v651_IBRION1
 vasp.6.5.1 10Mar25 (build Aug  9 2025 08:30:21) complex                         

   IBRION =      1    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
   9 F= -.57739597E+03 E0= -.57739604E+03  d E =-.258116E-03  mag=   164.2198
  10 F= -.57739615E+03 E0= -.57739627E+03  d E =-.182489E-03  mag=   164.2138
  11 F= -.57739618E+03 E0= -.57739629E+03  d E =-.310322E-04  mag=   164.2047
 reached required accuracy - stopping structural energy minimisation

------
v651_module
 vasp.6.5.1 10Mar25 (build Aug 12 2025 09:40:04) complex                         

   IBRION =      2    ionic relax: 0-MD 1-quasi-New 2-CG
   IALGO  =     68    algorithm
   8 F= -.57739465E+03 E0= -.57739471E+03  d E =-.465436E-03  mag=   164.2392
   9 F= -.55153539E+03 E0= -.55153536E+03  d E =0.258588E+02  mag=   164.2091
  10 F= -.57708801E+03 E0= -.57708807E+03  d E =0.306178E+00  mag=   164.2092
 reached required accuracy - stopping structural energy minimisation
 

[manuel_engel1]

Hi David,

Thanks for the bug report. It's indeed very strange to see these results and I'm not yet sure what the reason is. On the surface level, there have not been any changes to the optimization routines going from VASP 6.4 to VASP 6.5. However, I cannot rule out that this is a bug. It will take some time to investigate this issue and reproduce all the results. What stands out is that the runs that produce the strange results have the smallest amount of ionic steps (at least compared to the other conjugate gradient results). This might hold some significance.

I will report back if I find anything.

[dholec]

Dear developers,

I would like to ask if you have any news regarding our proposed issue. Thanks,

David

[manuel_engel1]

Dear David,

I have successfully reproduced some of your results, but due to the size of these calculations, it takes a long time. In fact, I am currently in the process of trying to find a smaller system where this can be reproduced. Do you have such a case or have you seen these symptoms in a more easily reproducible example?

[manuel_engel1]

If possible, I would also like to see the OSZICAR files of the individual runs.

[dholec]

Dear Manuel,

Many thanks for your efforts. I do not have any other/simpler case at this point (we reverted to v6.4.3 since then). The OSZICARs from the test runs I reported originally, are attached.

David

[manuel_engel1]

Dear David,

Here is a quick update. The jump in energy happens when switching from the blocked Davidson (DAV) algorithm to the RMM-DIIS algorithm. This is consistent across all your calculations. I have started to investigate this issue on an auxiliary system, but so far it's not reproducible. What you have found is probably more related to the stability of RMM-DIIS than to IBRION = 2. My guess is that rerunning your system with ALGO = normal would make these problems go away.