A puzzle for the 2nd ionic step

[yhan_vasp]

Dear VASP developer,

For many years, I have puzzled by the 2nd ionic step when I run VASP. I don't understand why the 2nd ionic step often (if not always) has a structure (and energy) so significantly deviating from those of the other steps (particularly of the neighboring 1st and 3rd steps).

The attached file "AlNO.rar" shows an example:
1st: energy without entropy= -53.68238070 energy(sigma->0) = -53.68238070
2nd: energy without entropy= -34.71051871 energy(sigma->0) = -34.71051871
3rd: energy without entropy= -55.13949860 energy(sigma->0) = -55.13949860
4th: energy without entropy= -55.68570517 energy(sigma->0) = -55.68570517
5th: energy without entropy= -55.90305439 energy(sigma->0) = -55.90305439
......
By checking "vasprun.xml", you should also be able to find that the structure of the 2nd significantly deviates from those of the other steps.

From my experience, this "2nd-ionic-step" puzzle will not influence the final results. However, I want to know what happen with such an "abnormal" behavior. Please help for the answer. Thank you very much for your time!

[martin.schlipf]

To understand this you need to be aware how the ionic minimizer works in VASP. Essentially, we obtain the next structure as a linear combination of the previous ones plus some displacements along the forces. Because in the second step only a single previous step is available the complete next structure is determined by the prefactor that you use to displace along the forces. For all other steps, we can already have a better starting guess, so that the forces are smaller.

You can investigate this by changing POTIM, which controls the size of the prefactor. When you reduce it, you should see smaller energy changes. Note that due to the nature of the optimization algorithms, structures with higher energies (or more precisely large forces) will have very little weight in the linear combination, so it doesn't affect the final result a lot. A too small step size can however lead to getting stuck in a local minimum.