Problem with DAVIDSON algorithm

[Michael B]

Hi,

I am currently running calculations on Ni(OH)2 and related Co(OH)2 
surfaces. Following a previously tested method I fix the spin of the system 
to a constant value rather than converging it. Since I ran into convergence 
problems for some of the systems I switched from the STANDARD 
diagonalisation scheme to the DAVIDSON algorithm. This resulted in a 
strange behaviour of the calculations:

*Problem I)*

Despite explicitly fixing the spin to a constant value the total spin is no 
longer constant but starts to change during the calculation when using a 
DAVIDSON algorithm. I.e. when fixing the multiplicity to 8 (7 unpaired 
electrons) I obtained after several geometry steps 10 unpaired electrons. 
(complete output file is attached as results.log)

When using the same input parameters apart from an increased smearing of 
0.3 eV (~3600 K) and a drastically reduced number of additional MOs (see 
2nd problem) in combination with the STANDARD algorithm the spin remains 
constant. The following commands were used to fix the spin (complete input 
attached as cp2k_input.inp):

---quote---
[...]
UKS TRUE
MULTIPLICITY 8
RELAX_MULTIPLICITY -1

[...]

FIXED_MAGNETIC_MOMENT 7

[...]
---quote---

Is this a planned behaviour and if so is there a possibility to fix the 
spin in combination with a DAVIDSON algorithm?

*Problem II)*

Using the DAVIDSON algorithm in combination with "SCF_GUESS ATOMIC" seems 
to result in a very strange convergence behaviour. Approximately 5 times 
the number of additional MOs are required (in the present case an increase 
from 20 ADDED_MOs to 100 ADDED_MOs was required) in order to avoid 
occupying all orbitals. Independent of whether one provides the required 
number of additional MOs or not each(!) SCF WAVEFUNCTION OPTIMIZATION step 
requires only 2 SCF steps to converge.

The obtained total energies for the single steps seem to vary randomly but 
the final total energy obtained for each geometry step remains after some 
changes during the first three steps constant. The obtained energies are 
off by approximately 20 Ha compared to what be expected based on previous 
results using a STANDRAD diagonalisation algorithm. The input strucutre was 
already reasonably well converged.

This problem is independent of the choice of system and the behaviour can 
reproducibly be switched on by the above combination of parameters. The 
relevant input and output files are attached to this message.

Interestingly replacing "SCF_GUESS ATOMIC" by "SCF_GUESS RANDOM" seems to 
cure this problem, i.e. the SCF cycle behaves normally and the obtained 
total energies are reasonably close to what one could expect based on 
previous calculations.

Is this behaviour due to a bug or simply due to an unlucky combination of 
parameters?

All the best

Michael
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20141029/f0919c74/attachment.htm>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: results.log.zip
Type: application/zip
Size: 2235434 bytes
Desc: not available
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20141029/f0919c74/attachment.zip>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: cp2k_input.inp
Type: chemical/x-gamess-input
Size: 9682 bytes
Desc: not available
URL: <https://lists.cp2k.org/archives/cp2k-user/attachments/20141029/f0919c74/attachment.inp>