Which ALGO tag is theoretically more accurate? I ran 2 geometry optimisations with IBRION=2 and ISIF=4, one with blocked Davidson and one with ALGO=fast, which begins with blocked Davidson and continues with RMM-DIIS.
The energies at the end came out identical, but the cell parameters and consequently the external pressure were slightly different (22.34 vs. 22.43kb).
My question is which scheme, if any, is theoretically more correct? I have done half of my calculations with the davidson algorithm, but the rmm-diis saved some time, so I want to know if changing to that for the rest, will give me a wrong overall answer.
Thank you!
ALGO-tag
Moderators: Global Moderator, Moderator
-
admin
- Administrator
- Posts: 2921
- Joined: Tue Aug 03, 2004 8:18 am
- License Nr.: 458
ALGO-tag
concerning all results related to the stress tensors, the most important
parameters to check are
--) the accuracy (PREC=Accurate, ENCUT= at least1.3* max(ENMAX)) of the run, in order to avoid aliasing errors.
--) the electronic an fore convergence parameters (EDIFF=xx, EDIFFG<0)
concerning the electronic optimisation algorithms: both algorithms should lead to the same converged results if all calculation's parameters are well converged. For large systems, the Davidosn algorithm has been found to be slightly more stable, but usually the pre-convergence is good enough to switch to RMM-DIIS after a few steps of blocked Davidson (ALGO = Fast)
parameters to check are
--) the accuracy (PREC=Accurate, ENCUT= at least1.3* max(ENMAX)) of the run, in order to avoid aliasing errors.
--) the electronic an fore convergence parameters (EDIFF=xx, EDIFFG<0)
concerning the electronic optimisation algorithms: both algorithms should lead to the same converged results if all calculation's parameters are well converged. For large systems, the Davidosn algorithm has been found to be slightly more stable, but usually the pre-convergence is good enough to switch to RMM-DIIS after a few steps of blocked Davidson (ALGO = Fast)
Last edited by admin on Mon Jan 30, 2006 12:57 pm, edited 1 time in total.
-
Tyx
ALGO-tag
Thanks for the reply, I guess I'll try the fast option. However, I was puzzled by your EDIFFG<0 comment... my EDIFFG is 10 times the EDIFF,, in my case 10^-6.If I understood correctly, in the manual it says that if negative it represents forces, so if I want my forces accurate to 10^-4, will setting EDIFFG=-1E-4 have the same effect and results as when it is positive? Is a negative setting faster?
<span class='smallblacktext'>[ Edited Mon Jan 30 2006, 02:36PM ]</span>
<span class='smallblacktext'>[ Edited Mon Jan 30 2006, 02:36PM ]</span>
Last edited by Tyx on Mon Jan 30, 2006 1:21 pm, edited 1 time in total.
-
admin
- Administrator
- Posts: 2921
- Joined: Tue Aug 03, 2004 8:18 am
- License Nr.: 458
ALGO-tag
no in general EDIFFG<0 is not faster, but especially for rather complex problems it is better to set an upper limit for the maximum remaining force than for a change in TOTEN with geometry, in order to avoid to get stuck in a (flat) local minimum.
(usually EDIFFG=-1e-2 to -1e-3 is a very reasonable choice)
(usually EDIFFG=-1e-2 to -1e-3 is a very reasonable choice)
Last edited by admin on Mon Jan 30, 2006 2:40 pm, edited 1 time in total.