Inconsistency between SCF and ICHARG=11 Calculations
Posted: Sat Mar 17, 2007 4:31 pm
The following may describe a feature rather than a bug, but it seems to present a significant problem. The issue is how well the CHGCAR file represents the charge and magnetization densities. I have noticed that the bands calculated in the SCF calculations differ significantly (band differences can change by 0.1eV) from those calculated using the CHGCAR from an SCF calculation and ICHARG=11.
To be specific about what I observe:
1. I run an scf calculation that generates a CHGCAR
2. I use that CHGCAR in a subsequent calculation, e.g. of bands along a symmetry direction (ICHARG=11).
3. I compare the band energies for the K-points in the EIGENVAL file that are common between the two runs, often the Gamma point is common.
4. I find that these band energies can be significantly different, usually the differences are a few hundreths of an eV, but I can provide cases where the differences exceed 0.1eV. It is not an overall shift since very often adjacent bands move in opposite directions.
I have noticed this for bcc Fe, fcc Cu and several other systems.
Is the internal real space mesh for the charge density finer than the one saved in CHGCAR? One test that I did was to run an scf calculation with a dense k-point mesh that generated a CHGCAR file. I then read that CHGCAR file into another non-self-consistent run with the same set of k-points (ICHARG=11) changing nothing else. The differnces of band energies were on the order of a few hundreths of an eV up to a tenth of an eV. However, if I allow another scf calculation (ICHARG=1) the band energy differences are much smaller (at least an order of magnitude smaller).
Here are some things that did not work -
i. CHGCAR is written with 11 digits. I recompiled the program to read and write with 14 digits - no change
ii. I increased the number of k-points - the eigenvalues all changed slightly but the inconsistency was hardly affected.
iii. I increased the precision from the default to high - the eigenvalues changed slightly but the inconsistency was not affected.
If anyone is interested in helping, I can provide input files, but I suspect you can observe this effect quite easily. The version of VASP used was 4.6.26. I also saw the same "feature" in version 4.6.28.
I would be grateful for any help or advice. If anyone runs a system as described in 1-4 above and does not notice discrepancies in the band energies please let me know - I hope that I am making a simple (or at least correctable) mistake.
Note added: March 19 2007
The inconsistency seems to be somewhat less for non-d systems. This led to the note in the guide that the LMAXMIX parameter can affect CHGCAR. Setting LMAXMIX = 4 significantly reduced, but did not eliminate the inconsistency.
Thanks, Bill Butler
<span class='smallblacktext'>[ Edited ]</span>
To be specific about what I observe:
1. I run an scf calculation that generates a CHGCAR
2. I use that CHGCAR in a subsequent calculation, e.g. of bands along a symmetry direction (ICHARG=11).
3. I compare the band energies for the K-points in the EIGENVAL file that are common between the two runs, often the Gamma point is common.
4. I find that these band energies can be significantly different, usually the differences are a few hundreths of an eV, but I can provide cases where the differences exceed 0.1eV. It is not an overall shift since very often adjacent bands move in opposite directions.
I have noticed this for bcc Fe, fcc Cu and several other systems.
Is the internal real space mesh for the charge density finer than the one saved in CHGCAR? One test that I did was to run an scf calculation with a dense k-point mesh that generated a CHGCAR file. I then read that CHGCAR file into another non-self-consistent run with the same set of k-points (ICHARG=11) changing nothing else. The differnces of band energies were on the order of a few hundreths of an eV up to a tenth of an eV. However, if I allow another scf calculation (ICHARG=1) the band energy differences are much smaller (at least an order of magnitude smaller).
Here are some things that did not work -
i. CHGCAR is written with 11 digits. I recompiled the program to read and write with 14 digits - no change
ii. I increased the number of k-points - the eigenvalues all changed slightly but the inconsistency was hardly affected.
iii. I increased the precision from the default to high - the eigenvalues changed slightly but the inconsistency was not affected.
If anyone is interested in helping, I can provide input files, but I suspect you can observe this effect quite easily. The version of VASP used was 4.6.26. I also saw the same "feature" in version 4.6.28.
I would be grateful for any help or advice. If anyone runs a system as described in 1-4 above and does not notice discrepancies in the band energies please let me know - I hope that I am making a simple (or at least correctable) mistake.
Note added: March 19 2007
The inconsistency seems to be somewhat less for non-d systems. This led to the note in the guide that the LMAXMIX parameter can affect CHGCAR. Setting LMAXMIX = 4 significantly reduced, but did not eliminate the inconsistency.
Thanks, Bill Butler
<span class='smallblacktext'>[ Edited ]</span>