Large atomic forces when LSORBIT = .True. with vasp6.1.0

[ttadano]

Hello.

When I tried to optimize the internal coordinate of a system with spin-orbit interaction (LSORBIT = .True.),
the atomic forces became very large with VASP 6.1.0. This phenomenon was not observed in VASP 5.4.4.

Before running an optimization with SOI, I relaxed the internal coordinate without SOI by VASP 6.1.0 and copied CONTCAR as POSCAR.
Therefore, the structure in POSCAR should be reasonable.

Below are the excerpts from the OUTCAR files.

VASP 5.4.4

Code: Select all

 POSITION                                       TOTAL-FORCE (eV/Angst)
 -----------------------------------------------------------------------------------
      1.33368      1.07700      0.29010         0.041951      0.000000      0.033320
      7.12718      1.07700      1.86290         0.041943     -0.000002     -0.033316
      4.35030      3.23100      2.52640         0.013505      0.000001     -0.002236
     10.14380      3.23100      3.93260         0.013508      0.000000      0.002235
     10.03160      1.07700      2.07089        -0.032839      0.000001      0.004507
      1.66492      3.23100      2.29925        -0.022614      0.000000     -0.027963
      7.45842      3.23100      4.15975        -0.022615      0.000000      0.027961
      4.23810      1.07700      0.08211        -0.032839     -0.000000     -0.004509
 -----------------------------------------------------------------------------------
    total drift:                               -0.003161     -0.000003      0.000006
 

VASP 6.1.0

Code: Select all

POSITION                                       TOTAL-FORCE (eV/Angst)
 -----------------------------------------------------------------------------------
      1.33368      1.07700      0.29010        12.832589      1.333413    -24.539761
      7.12718      1.07700      1.86290       111.850117    -24.971320    129.926731
      4.35030      3.23100      2.52640      -121.755018     -1.718684    -67.081141
     10.14380      3.23100      3.93260      -121.078815     -2.751189    -35.697358
     10.03160      1.07700      2.07089       -11.748897      0.129917   -120.542707
      1.66492      3.23100      2.29925       -92.867240     -1.443072    -40.272766
      7.45842      3.23100      4.15975        52.805601     30.196516    280.629710
      4.23810      1.07700      0.08211       169.961663     -0.775582   -122.422708
 -----------------------------------------------------------------------------------
    total drift:                              841.945960      9.049849    410.561481
 

The total energy and stress tensors are almost the same between VASP 5.4.4 and VASP 6.1.0.
The large difference was observed only for atomic forces, particularly in the non-local components.

I have attached the relevant input and output files for your consideration.
Any help or comments would be highly appreciated.

Best regards,
Terumasa

[ferenc_karsai]

That is indeed a bug.
Thank you very much for bringing it to our attention.

In the course of the refactoring we did to port VASP.6 to GPUs by means of OpenACC we unfortunately introduced a bug.
This bug affects the forces in non-collinear calculations with NCORE /= 1 (as in your job) with VASP.6.1.0.

This bug is now fixed. The fixed version is uploaded to the download portal (VASP.6.1.1).
Please use that version in future.

[ttadano]

Thank you for your feedback and the quick fix.
I've tested with VASP.6.1.1 and confirmed the issue has been solved.