The ABINIT Users Mailing List ( CLOSED )

["Zhang Mike" <mz24cn@hotmail.com>]

Dear Verstraete,

Thanks for your helpful comments. Here is my brief reply. 

> Hello Mike,
>
> a few comments, as I've also done a few surface calculations. 
>
> You should try optcell 9: this keeps the c axis fixed (and the vacuum 
> layer between your slabs), while allowing an in-plane relaxation.
I think "in-plane relaxation" may be not suitable for my occasion, so I 
didn't ever try it.

> Using ionmov 2 (you don't need moldyn, it's less efficient) will relax
> your atomic positions, but the problem can be that you want a surface
> reconstruction. In this case you need a supercell, and you need to try
> breaking the symmetry by hand, and letting abinit find a new minimum
> itself (or guided by you if you know what you're looking for). I don't
> think the calculations we've done so far are questionable, simply if we
> haven't allowed for surface reconstruction they may not give the
> experimental result, or the absolute minimum. The advantages of enforcing
> symmetry in abinit are incommensurable with the small difficulty of
> breaking symmetry by hand to see if things reconstruct.
Yes. Just because I used ionmov=2, my calculations were converged. I found 
ionmov=2 might often work when ionmov=3 failed to converge the calculation. 
Because ABINIT can not realize the symmetry breaking, I modified the 
surface atoms to slightly depart the equilibrium positions by hand, just as 
you suggested, and got the right result. (I have reported it in my newly 
accepted manuscript) If I didn't modified the surface atom positions, I 
also could converge the calculation with the semicore pseudopotential 
(failed with nonsemicore PP), though it was very difficult and the 
res-forces were unstable, and got the wrong symmetry-preserving result.

> The C-P algorithm is great, but I don't think it is a solution to your
> problems. The convergence difficulties in surface calculations, in my
> experience, are self-consistent cycle problems, not geometrical relaxation
> ones. In the latter context CP would be nice, but the problem is inherent
> in the vacuum. The potential there can fluctuate wildly without
> influencing the energy, so the convergence is slow (sometimes impossible).
> CP would probably have the same problems. A clean-up of the SCF might be
> needed, but a more efficient or adapted preconditioning scheme would be
> even better (anyone heard of one?).
I ever encountered the self-consistent cycle problems, as you proposed, but 
I think we should not confuse it with the structural relaxation problem. 
Now in my calculations, SCF cycle can be finished in 30-50 steps (by 
properly setuping the surface geometry, choosing the k points, the PP, and 
iscf), while the res-force seems very difficult to converge. Even it 
reached the convergence criterion (I often use tolmxf=5.0d-5), but the 
trend of tolmxf with the steps is oscillatory, seems not to converge. We 
can see even SCF convergence was successful, the relaxation convergence was 
not acquired yet. Even the relaxation convergence was acquired, it may 
still converge to a wrong symmetry-preserving results. Thus I think the 
difficulty to converge the relaxation is much associated with the algorithm 
ABINIT applied, not much related with SCF convergence. I think C-P 
algorithm is effective, especially for the occasion of many-atoms 
relaxation. 

> To boot, the smaller your gap the smaller your CP time step has to be
> (hence it could even be slower than Broyden minimization). Schemes exist
> to work with metallic systems (VandeVondele & DeVita), but none of this
> has been implemented widely by other groups, so your slabs would have to 
> be insulating to begin with... I dread the thought: all of my slabs thus 
hoho, Yes, as you guessed, I am conducting insulators/semiconductors 
calculations. Many people in the forum seem using ABINIT to study the 
insulators/semiconductors, Maybe ABINIT is not suitable for 
insulators/semiconductors yet? :(

> far have been metallic.
>
> If anyone else has thoughts on the SC convergence in abinit, I'm 
> interested.
As described above, I found properly setuping the surface geometry, 
choosing the k points, the PP, and iscf, die***, seems conducive to the SCF 
convergence.

> Ciao
>
> Matthieu

Mike Zhang





_________________________________________________________________
享用世界上最大的电子邮件系统― MSN Hotmail。  http://www.hotmail.com