Slow convergence

[forsdan]

Dear all,

I am experiencing a common problem with slow electronic convergence. My system consists of an interface with five layers NbN on top of five layers Fe (15 atoms total). I'm trying to carry out a number of convergence tests, but the required number of steps exceed >60, but each calculation converge before 100 steps.

However, I would like to speed up the convergence, so I switch to a linear mixing (AMIX = 0.1-0.2 , BMIX = 0.0001, AMIX_MAG = 0.8, BMIX_MAG = 0.0001) but with no improvement. Then I tried to increase BMIX (since GAMMA was smaller than 1) to 2.0-3.0, while keeping the default settings. No improvement there either. I also switched the ALGO-tag to F, and combined them with the different mixing.

I also tried the above procedure for different k-point sampling, but with no reduction of the number of steps. Hence calculations at a sufficient k-point mesh will be quite time-consuming. Is there something else I may try?


Best regards
Dan Fors

INCAR (without mixing tags)
----------------------------------------------

SYSTEM = FeNbN interface
ISTART = 0
ICHARG = 2
EDIFF = 1E-5
ENCUT = 500 eV
ENAUG = 520 eV
PREC = Accurate
LREAL = .FALSE.
ISMEAR = 1
SIGMA = 0.05
NSW = 0
NELM = 200
VOSKOWN = 1
ISPIN = 2
MAGMOM = 5*0 5*0 5*3
LWAVE = .FALSE.
LCHARG = .FALSE.

POSCAR
--------------------------------------
FeNbN interface
1.0000000000000000
3.0739999999999998 0.0000000000000000 0.0000000000000000
0.0000000000000000 3.0739999999999998 0.0000000000000000
0.0000000000000000 0.0000000000000000 24.7645999999999979
5 5 5
Direct
0.5000000000000000 0.5000000000000000 0.0000000000000000
0.0000000000000000 0.0000000000000000 0.1292530466876105
0.5000000000000000 0.5000000000000000 0.2585060933752211
0.0000000000000000 0.0000000000000000 0.3877591400628316
0.5000000000000000 0.5000000000000000 0.5170121867504421
0.0000000000000000 0.0000000000000000 0.0000000000000000
0.5000000000000000 0.5000000000000000 0.1292530466876105
0.0000000000000000 0.0000000000000000 0.2585060933752211
0.5000000000000000 0.5000000000000000 0.3877591400628316
0.0000000000000000 0.0000000000000000 0.5170121867504421
0.0000000000000000 0.0000000000000000 0.6543049352705070
0.5000000000000000 0.5000000000000000 0.7064055143228640
0.0000000000000000 0.0000000000000000 0.7585060933752211
0.5000000000000000 0.5000000000000000 0.8106066724275781
0.0000000000000000 0.0000000000000000 0.8627072514799350

[lunmei]

try this:
1) use davidson (Algo=N)
2) if method 1) do not work ,
set ICHARG = 12, if it fails to converge, report the bug.
if ICHARG = 12 converges, try ICHARG = 2, with AMIX= 0.1, BMIX = 0.01.
if not work, try, ICHARG= 2, with BMIX= 3.0, AMIX=0.01.

I got this information from previous topic, and it works fine for my system.

[forsdan]

Dear all,

I'm have to come back to this issue once again, because i'm still struggeling with the problem. I've made som minor changes to the INCAR above, (Added LMAXMIX = 6, set SIGMA = 0.02 and increased MAGMOM to 5*3 5*0 5*4).

However, I've now tried the mentioned procedure. I also tried starting with linear mixing for a different number of initial steps (8-15) and then continued with the default settings. The number of self consistency runs still exceed 100 iterations, even 200 at some interface distances.

I've check the positions as well and it all looks reasonable. If there is any further suggestions about what I may try I would appreciate it.

Best regards,
Dan Fors

[mssb2]

Dear forsdan,
Have you sollved your problem regarding the slow electronic convergence??
I am doing calculation on Fe cluster, & facing the same convergence problem. I switched on the lfollowing tag:
AMIX=0.2
BMIX=0.0001
AMIX_MAG=0.8
BMIX_MAG=0.0001
But there is no improvement in convergence.
I tried for different combination of the tag.
It will be appreciable if you have any suggesstion to speed up my calculations.
Thanking you in advance..

[forsdan]

I'm afraid I still havn't figured out this particular problem yet. I have so far only encountered it for the interface systems Fe/NbN and Fe/NbC, while the bulk systems work fine.

The slab/vacuum systems have generally slow convergence as well, but still converges reasonable fast for me (< 70 electronic steps), which I can live with for the moment, so I havn't given them more thought.

However, if someone has suggestions for other mixing-tags settings or ways of improving relaxation for the interface systems, I would appreciate to hear them.

Best regards
/Dan Fors


<span class='smallblacktext'>[ Edited Mon Apr 09 2007, 02:55PM ]</span>