SCF procedure does not converge during ML-FF MD training trajectory

[julien_steffen]

Dear VASP community,

we are currently trying to learn a ML-FF for a Ga SCALMS system with additional H atoms. To that reason, we placed five H2 molecules in vacuum over a Ga liquid (see POSCAR).

During the ML-FF MD calculation the very first SCF cycle is terminated with the error message Error EDDDAV: Call to ZHEGV failed. Returncode = 15 1 16 after a few SCF steps. Looking at the energy change etc., it could bee seen, that the cycle does not convergence properly.

We decided to perform the calculation with ISPIN = 2 based on previous ML-FF calculations, where the trajectory will be more stable, if the system is formally spin-polarized. However, for this system we already tried to change ISPIN to 1 (see directory no_spin). Then, the SCF cycles convergence properly. However, looking at the XDATCAR, the trajectory seems not plausible. H2 molecules are splitted and all molecules move too fast.

Since the file size for the folder no_spin is exceeds the maximum filesize per attachment, the corresponding files are uploaded to our cloud system:
https://faubox.rrze.uni-erlangen.de/get ... TN/no_spin

Best regards

[alex]

Hi julien_steffen,

sometimes it helps to change simply the SCF algorithm. I'd suggest to go 1st for ALGO = V and try again, no changes otherwise.

2nd option: try a different VASP exe with another linear algebra package.

Good luck!

alex

[ferenc_karsai]

I looked at your structures from the XDATCAR and I also observed that some of the H atoms diffuse into the gallium bulk. In the last structure one H atom comes suspiciously close to one Ga atom.

Although you have a time step of 1fs, it is maybe too large for the regular hydrogen mass of 1. I would suggest to increase the hydrogen mass between 4-8 times (first try 8 and later you can lower) in the POTCAR file or the INCAR file (change POMASS). We have this also written on our best-practises site under "Molecular dynamics setup":
https://www.vasp.at/wiki/index.php/Best ... rce_fields