SCAN+rVV10 Subspace-Matrix Non Hermetian

[mwuensch]

Hey,
I've encountered a problem quite a few times now when relaxing a slab structure using SCAN+rVV10 with VASP 6.3.2.
At the point where I assume the wavefunction would be written instead I get hundreds of lines saying "WARNING: Sub-Space-Matrix is not hermitian in DAV". I looked up what might be the cause of this, but didn't find a satisfying answer, other than maybe the VSC's LAPACK installation has an issue. I will try compiling my own VASP version on the cluster in the near future, but for now I wanted to ask here if there maybe is another reason for my problem.

All the input and output files are attached. Run 1 was done using PBE_sol, run 2 using SCAN+rVV10. The input files are those for the SCAN run, so the POSCAR is the result of the previous PS run, while the CONTCAR is the result of the SCAN run (CHG, CHGCAR & WAVECAR excluded due to size). I'm also curious about the discrepancy between the calculated energies, SCAN seems to consistently yield about 2x the energy of PS, which seems a little beyond coincidence. SCAN: F= -.10687547E+04; PS: F= -.53232435E+03.

Any adivce or help is appreciated!
Cheers
Max

[alex]

Hi Max,

I usally set NPAR instead of NCORE (old habit) to sth. small like 4 to 8. If your error encounters, I'll double it until the error disappears.
However, in my cases the numerical mess starts immediately. Lucky me. :-)

Cheers,

alex

[alexey.tal]

Dear Max,

I have not been able to reproduce this issue so far. I see that there is a jump in the rVV10 correction energy in the last iteration:

Code: Select all

Total rVV10 correction in eV:     14.0275068   776.6849543

Code: Select all

Total rVV10 correction in eV:     38.1678384  1734.6968325

and these values in the last iteration seem to be incorrect.

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

Often such error occurs when the is a problem with the geometry. In your calculation that doesn't seem to be the case.
Also, this error only occurs in the blocked-Davidson algorithm, so you can switch to RMM-DIIS (ALGO=VeryFAST) to avoid it.
Could you please show the other calculation where you've encountered this problem?

I'm also curious about the discrepancy between the calculated energies, SCAN seems to consistently yield about 2x the energy of PS, which seems a little beyond coincidence. SCAN: F= -.10687547E+04; PS: F= -.53232435E+03.

Comparing the total energies calculated with different xc functional is not very meaningful from a physics point of view.

[mwuensch]

Thanks for the replies!

To Alex:
I'm currently running another calculation with NCORE = 1 (/NPAR = max), I somehow doubt that'll fix it, but who knows! Thanks for the tip!

To Alexey:
I'm attaching another two calculations where I've done one PS run and 2 SCAN runs/2PS and then one SCAN run. Structure-wise they only differ by an addition of some Hydrogen atoms, or a vacancy on the surface, hope that makes the data different enough to be relevant to you. The outputs look the same to me.
I'll be sure to use IBRION = 1 for another run, thanks for the tip!
Are you unable to reproduce the issue on VSC5's 6.3.2. module installation or on your own local one? If it's on your local one I guess that's my queue to compile VASP myself for future calcs.

Cheers
Max

[alexey.tal]

Are you unable to reproduce the issue on VSC5's 6.3.2. module installation or on your own local one? If it's on your local one I guess that's my queue to compile VASP myself for future calcs.

I ran your calculation on our local machine, but I had to reduce the number of k-points, otherwise, the calculation is too large.
Also, could you run one of these calculations once again to see if the error is reproducible or if it occurs sporadically.

[mwuensch]

I've now run around 20 calculations using INCAR settings with IBRION = 2 and SCAN+rVV10. Regardless of NCORE setting or energy cutoff (possibly reducing KPOINTS might yield other results, currently trying 3 3 1 instead of the 9 9 1 I've tended to use, will keep you posted) the same error appears. I'm attaching a folder that contains the output of 3 separate relaxation runs that all look the same.

[mwuensch]

The 3 3 1 k-points run is done and also ended in the same error, I'll attach those files as well

[alexey.tal]

Thank you for the extensive testing that you've done. It might be a bug in the code and we are going to look into it.

Meanwhile, I have two suggestions that you could try:
1) Your convergence criteria are too stringent. You should set EDIFFG=-0.001 or even less stringent. The problem is that SCAN functional leads to noisy forces and it is practically impossible to achieve a better convergence.
2) IBRION=2 is more sensitive to the noisy forces, so you should try IBRION=3 (damped molecular dynamics) instead, which should behave better in this case.

[mwuensch]

Thank you for your advice! Is the EDIFF setting of 10^-5/6 appropriate or could I reduce that with minimal consequence as well?

[alexey.tal]

Since your issue isn't related to the convergence of the electronic iterations, you don't need to change EDIFF.