Which approach is better for structure relaxation (ISIF=3)?

[alpinnovianus]

As titled,

suppose I have a tetragonal structure and I want to relax it using ISIF=3, EDIFF, EDIFFG to get the lattice parameters a,b, and c.

I notice that I can arrive at similar relaxed lattice constants a, b, c from two approaches, differing in whether I manually displace one of the atoms in my POSCAR for initial configuration, e.g.:

Direct
0.000000000 0.500000000 0.250000000
0.000000000 0.500000000 0.750000000
0.500000000 0.000000000 0.250000000
0.500000000 0.000000000 0.750000000
0.000000000 0.000000000 0.000000000
0.500000000 0.500000000 0.000000000
0.500000000 0.500000000 0.500000000
0.000000000 0.000000000 0.500000000
0.750000000 0.250000000 0.000000000
0.250000000 0.750000000 0.000000000
0.750000000 0.750000000 0.000000000
0.250000000 0.250000000 0.000000000
0.250000000 0.750000000 0.500000000
0.750000000 0.249000000 0.500000000
0.251000000 0.250000000 0.500000000
0.750000000 0.750000000 0.501000000

or to keep them in the symmetry positions without displacing them (0.25, 0.5, or 0.75).

After the calculations finish, the manually displaced structure will have CONTCAR that isn't tetragonal, i.e. we have the off-diagonal terms:

5.5246022033677544 0.0000052421620465 -0.0000061392476270
0.0000068317938047 5.5245558229953451 0.0000040301375807
-0.0000078393243090 0.0000050551768983 6.8885820111049139

and the relative coordinates are all moved:

Direct
0.0000591313384665 0.4999394057025163 0.2500766918597011
0.0000960386480647 0.4999031528527850 0.7500595431593108
0.5000942204267217 -0.0000947125196263 0.2500580912468196
0.5000585977256463 -0.0000590355203524 0.7500770259952698
0.0000383719367016 -0.0000232861815960 0.0000281190030009
0.5000370986808886 0.4999721503457955 0.0000282936164365
0.5000277955163501 0.4999600546046230 0.5000319276268052
0.0000273178312206 -0.0000353781058559 0.5000322823391325
0.7500706088558012 0.2499561607663021 0.0000316118232972
0.2500082757329685 0.7499860627982675 0.0000312847603286
0.7500968749870348 0.7499604661449809 0.0001343392310228
0.2500275084558400 0.2499197926518050 0.0000896858522983
0.2501013904963621 0.7498972304456063 0.4999908423747649
0.7500363721281671 0.2499555592350028 0.4999905956886319
0.2501809635571148 0.2499475749177189 0.5000029045010441
0.7500394336826489 0.7498148018620299 0.5003367609221337

The diagonal terms are very similar to the values obtained without manual displacement.
The forces are finite but I can see they are below EDIFFG.

So with the manual displacement we lose the tetragonal symmetry, and this brings up further questions like the choice for high-symmetry points for bandstructure if the symmetry is always lost, etc.

but if I don't apply the manual displacement,

sometimes I get a relaxed structure that has the same resulting a, b, and c lattice constants as the diagonal terms above, retain its tetragonal symmetry, but the relative coordinates aren't changing from their symmetrical positions and the forces are all shown as zero in the OUTCAR.. This last part is what concerns me, if the lattice constants a, b, and c, are the only parameters that change, whether or not this is still a "structure relaxation".

I wonder which of the two methods are the preferred/correct one?

• apply manual displacement and lose the symmetry

• don't apply manual displacement and keep the symmetry

assuming that the total energy and the a,b,c lattice constants from the two methods can be considered very similar.

Thank you!

[alex]

Hi alpinnovianus,

well, as you already figured out: both results are comparable. The question you may wish to answer: 'why they are not identical?'

a) symmetry routines give precise positions wrt. space groups, so you'll end up with exact Wykhoff positions even after structure optimization.
b) we are doing numeric approximations here. depending on your settings in the INCAR (e.g. EDIFFG) the convergence is reached, but machine precision goes beyond that. So you end up with sth. like 0.00000rubbish at the last digits. This is random.
As you pointed out the results are the same. I would prefer the symmetrized solution if it makes sense in a scientific point of view.

Happy VASPing! :-)

alex