My Community

Dear VASP Support Team,

I am using VASP 5.4.4 version to run Molecular Dynamics of a Pt8 cluster in gas phase, also Pt8 with several atoms attached. Simulations are running on a supercomputing center which allows me to paralellize the calculations.

Although total number of steps (NSW = 2000), the simulation runs until some certain "random" step when it eventually crashes, otherwise it will finish the job normally.

For me is a 'random ' step because for different isomers of Pt8 with same INCAR and slurm_script one crashes at (i.e) 1595 Steps another one at 30 Step and another one at 644 Steps while sometimes, the calculation finish normally, and i restart from CONTCAR to get the next 2000 steps.

After restarting, it can finish calculation or stop at some random step as mentioned before.

So, when i look into the and out.#### file a long tail output message is shown repeatedly:

{ 4, 6}: On entry to
PDSTEIN parameter number 4 had an illegal value
{ 23, 4}: On entry to
PDSTEIN parameter number 4 had an illegal value

I have also searched in the Vasp community Forum and found this answer by merzuk.kaltak that speaks about mixing parameters.

I don't know how to play wth AMIX and BMIX so I check the Vasp wiki and found that MAXMIX parameter is the one ruining my simulations. So i decided to eliminate it form INCAR file, and now seems that simulations are ending normally, despite longer calculation times.

I am posting this in order to learn why that error was arising and also how to tune the mixer parameters IMIX, AMIX, BMIX, and MAXMIX to perform efficient calculations without crashes.

Any help or advise would be extremely helpful so thank you very much in advance for your reply.

If any file is missing or further explanations are needed please do not hesitate in contact me, I would gladly give you any requested info.

Thank you very much,


Ramon

I have a few comments/suggestions that you could consider instead of trying to change the mixing parameters.
1. You have a small unit cell of 8 atoms in which case the reciprocal space projection scheme is more efficient (LREAL=.FALSE.)
2. The system you are considering seems to be metallic so you should use an adequate ISMEAR (1 or -1)
wiki/index.php/ISMEAR
3. Since it is such a small lattice using so many CPUs (192) with your current settings might not be particularly efficient: note that VASP is computing a lot of empty bands which is unnecessary for MD runs. You might run the calculation in a similar time using for example 48 cores only (do your own testing). Setting NCORE>1 (wiki/index.php/NCORE) might help you to have good scaling for 192 cores but again do your own testing.
4. Have a look at the structure at the different MD steps to see if everything is still according to what you would expect (no overlapping atoms, breakdown of the cluster of atoms etc). The electronic self-consistency seems to be working correctly until an ionic step where it does not converge anymore (more than 60 SCF cycles) and after that, I think the forces might be inaccurate and the updated structure in the next step will be wrong.
5. If none of these suggestions helps, then you might also consider changing ALGO=Fast to ALGO=Normal.