VASP Forum

Dear experts,
I am running simulations of an hcp Ti supercell with 150 atoms where 2 atoms are replaced by other species.
The simulations run at the Zuse-Institute Berlin on the cluster there.

For one of the runs, after the calculation seems to converge, the energy gradually increases again with the ions moving to strange (=absurd) positions. VASP finally quits with the error message
Error EDDDAV: Call to ZHEGV failed. Returncode = 27 2 48
but at this point the ionic positions are completely off.

I did run the same job again without changes; weirdly, the job now completes without problems.
The staff from ZiB says that this being due to some hardware or other machine issue is very unlikely, so I suspect that some weird round-off issues or so may be responsible.
I've tried to analyse the results (with my limited knowledge of the inner workings of VASP) and found that the forces on the ions (especially ion 0) become different at the moment when the results start to differ. The detailed force output mainly differs in the "convergence-correction"-term.

I've attached the relevant output (OUTCAR, slurm etc); the files with the extension -8695526 belong to the job that failed, the ones without to the second, converging job.

Thanks for any insights to this.
Best regards,
Martin.

Dear Martin,

Thanks for reaching out to us on the VASP forum. Me and my colleagues have looked at your calculation. It is not yet fully clear what is causing the issue. However, it is unlikely that it has something to do with the conjugate gradient algorithm during ionic minimization. Rather, it seems that there is a problem with the electronic minimization at some point that escalates and causes the error.

Here are some suggestions based on our current understanding of this issue:

• Increase the smearing SIGMA. Since this system is metallic, too small values can cause convergence issues.

• Set ISPIN=1. It appears that this system is not magnetic, so convergence will be much faster and also more robust without an additional spin degree of freedom.

If I find any additional insights, then I will update this topic. Similarly, if you find that your calculation is still not converging consistently, feel free to reach out again.