UT theoretical chemistry code forum

Hi,

I've been trying bader analysis out and tested several species. The results are somewhat bizarre. All calculation were done using VASP (vasp.4.6.31 08Feb07) with PAW-LDA following the instructions on http://theory.cm.utexas.edu/vasp/bader/vasp.php. Basically stating, 10 A box, with two atoms in the middle with an initial distance and relax once. All forces on atoms reached limit. The bader analyses were done at final step statically.

Here is a list of elements I tested.

element ///// length ///// transfer

F2 ///// 1.3986402064513740 ///// 0.0409
Cl2 ///// 1.9788965537567780 ///// 0.0095
Br2 ///// 2.2748185297021860 ///// 0.0189
I2 ///// 2.6460187853865310 ///// 0.0263
Na2 ///// 2.9926281221607560 ///// 0.0164
K2 ///// 3.8682669985132410 ///// 0.0100
C2 ///// 1.2623078710528580 ///// 0.2370
Si2 ///// 2.1890866403177670 ///// 0.0452

BrCl ///// 2.1292623432542480 ///// 0.1453
BrF ///// 1.7637956410893430 ///// 0.4469
H2 ///// 0.767 ///// 0.0155

The charge transfer got from ACF.dat seem to be off a bit generally, with some big offs.

I was wondering if my procedure has some systematic mistakes or VASP cannot be that precise about 10^-2.

Thanks!
Terence

Make sure to look for convergence with respect to the FFT grid density.

[quote="graeme"]Make sure to look for convergence with respect to the FFT grid density.[/quote]
Thanks for the reply. I checked the total charge bader detected, all integers equal to the number provided by the potential. The density-grid output by bader is about 70 x 70 x 70 to 100 x 100 x 100. I'm not sure if they are dense enough.

I increased the FFT grid to 150 150 150 (for Br2 and I2), 210 210 210 (Cl2 and F2) and the charge transfer went below 0.01! Thank you! I didn't expect such large values.

Everything else is fine but for carbon-carbon pairs the charge transfer is still not under 0.01. I have tried NGX(,Y,Z)F from 220 to 400, PAW-LDA and PAW-GGA potential C, C_s and C_h. The two kinds of potentials behave very alike in terms of this issue, but C and C_s/C_h behave differently. The transfer is

C_LDA:
NGXF 220: 0.2677
C_s_LDA:
NGXF 220: 0.0175; NGXF 400: 0.0154
C_h_LDA:
NGXF 220: 0.0176; NGXF 300: 0.0240

C_GGA:
NGXF 220: 0.2870
C_s_GGA:
NGXF 220: 0.0175
C_h_GGA:
NGXF 220: 0.0177

It seems that C_s and C_h give better results but can't reach below 0.01. I also checked the CHGCAR plot in VESTA, which showed the split surface a little favoring one atom, but the surface is flat and smooth.

Planar dividing surfaces are somewhat pathological since they can align with the grid. Basically, you won't get nice cancelation of errors as compared to a curved surface or one that is not aligned with the grid because any systematic bias at one point on the surface is the same as the one next to it. We are implementing a better interpolation scheme, but for now, we have to live with slow convergence as a function of grid density.

A simple check for this, by the way, is to rotate your molecule by any random angle. Likely, the Bader charges will become more equal for the two atoms.