Strange problem from running VASP

[yhan_vasp]

Dear VASP developer,

I tested two VASP jobs, as attached ISIF3-1 and ISIF3.

ISIF3-1 is the original job that I ran. You can find that the converged energy is -1706.85702607 eV. Then, I copied ISIF3-1 to ISIF3, and only changed CONTCAR to POSCAR. I expected that two jobs should have the same energy.
However, after I ran this job, I got -1706.83425953 eV. The large energy difference of > 20 meV does not make any sense to me at all.

How do you explain this strange (unreasonable) result?

Thank you for your help and time.

[martin.schlipf]

You calculate a rather large system. For most cases, the meaningful value is the energy difference per atom or formula unit. So your error of ~0.1meV/atom seems reasonable. Perhaps only if you want defect formation energies a higher precision would be required.

[yhan_vasp]

So, such error depends on the initial configuration (POSCAR), because the only difference of two jobs is POSCAR. Is it possible to avoid or reduce this error from a more accurate way?

[martin.schlipf]

You can set a smaller EDIFF to see if you obtain the same results. But please be aware that the more complex your system/problem gets the more likely you can end up in a local minimum. So your results will depend on the starting conditions. If you really need such a high precision the only way is to study a few different configurations and take the result with the lowest energy.

[yhan_vasp]

I reran the 2nd job with EDIFF=10^-8, but the energy has no change.

Because I copied the CONTCAR from the 1st job to be POSCAR for the 2nd job, the fully relaxed configuration of the 1st job is the same as the initial and final configurations of the 2nd job (note that the 2nd job converges within only one ionic step without any relaxation). So, logically, these two configurations are completely the same and then should have the exactly same energy, but the energy of the 2nd job increases ~20 meV. For such situation, I judge that, very likely, the optimization algorithm for ISIF = 3 should be problematic.

[martin.schlipf]

Sorry my bad, I had completely missed that you were running volume relaxation calculations. Please consider the following section on the wiki. In a nutshell, the plane-wave basis is a function of the lattice vectors but fixed during the relaxation. However when you restart the calculation the basis is reset and may be different than in the original run. In fact when you check your OUTCAR files and look for the number of plane waves, you'll see
maximum number of plane-waves: 76482
maximum number of plane-waves: 76716
so the basis is different and you would not expect the same total energy. The difference should decrease with larger ENCUT, though.