total energy increases during geometry optimization

[mindflying]

Dear all,

I have a ZnO slab of 14 layers of atoms, with one O vacancy on either surface layer, and the 2 central layers fixed. The structure was pre-optimized using normal precision and conjugate-gradient method so that forces are below 0.15 eV/angstrom for unfixed atoms. This part went ok, and the total drift is below 0.2 eV/angstrom.
Then I started optimizing the pre-optimized structure using high precision and quasi-newton method. At this stage, for the first 30 ionic steps the total energy went down sharply, and then gradually went up for the next 50 steps (before it found the convergence) although the energy increased was very small, only 0.0002 eV. I set the convergence criterion as force below 5 meV/angstrom. Another problem with high precision optimization is that the total force along x direction is larger than 0.4 eV/angstrom from the very first ionic step.

INCAR file for high precision:
ISMEAR = 0
SIGMA = 0.05
NPAR = 8
LREAL= Auto
PREC= HIGH # Normal for normal precision
ENCUT = 500 # 400 for normal precision
NELMIN = 8
NSW = 300
IBRION = 1 # 2 for normal precision
EDIFFG = -0.005 # -0.15 for normal precision
LCHARG = .FALSE.
LWAVE = .FALSE.

KPOINTS file:
Automatic mesh
0
gamma
2 3 1
0. 0. 0.

POSCAR (initial part):
zno_wz_100_face_clean
1.00000000000000
10.6010743242120604 0.0000000000000000 0.0000000000000000
0.0000000000000000 6.5724016831842542 0.0000000000000000
0.0000000000000000 0.0000000000000000 38.0243628917455254
O Zn
54 56
Selective dynamics
Direct
0.1915281007363206 0.0005947140714142 0.6502410214432612 T T T
...

POTCAR used was PAW_PBE.

Thanks for your help.

[alex]

QN is also good for climbing uphill and finding local maxima. Check your structure with a frequency analysis for imaginary modes.

Cheers,

alex

[mindflying]

Hi, Alex,

It looks so, at least in my case. Does this implies QN (quasi-newton) is not properly implemented in VASP as it first decrease the energy and then increase it? To check the imaginary modes is a good idea, although it's a different story.

[alex]

All optimisation algorithms are designed to find local stationary points. It will _not_ notice, if this is a minimum or local maximum.
You may wish to think about MD related stuff like annealing, if you want to avoid frequency analysis.
I'd stick to the first approach and play a little with switches. :-)

Cheers,

alex

[mindflying]

Hi, Alex,

While your point on optimisation is in general right, I think one should get a local minimum in the special case of structure optimisation if the algorithm is implemented properly. Since atoms are moved along force direction (for QN case), which is negative first derivative of energy, one expect the energy to go down if the movement is small enough (so not over shoot).

I also tried conjugate gradient method. It didn't help. Maybe it's good to check the modes at the final higher-energy configuration and at the minimum-energy configuration.

Thanks.

[alex]

Good luck!

In case this fails, too, you might wish to consider er Tomas Bucko's GADGET kit, which allows eigenmode following and very advanced optimisation stuff. It's great! But I'm not sure how far the implementation has proceeded. I was once using the stand alone version, which is great

Cheers,

alex

[mindflying]

I checked the max forces and found they just fluctuate, which reminds me maybe the target force is set too small (5 meV/A). I will set it to 10 meV/A and try.