[Pw_forum] GGA+U: Sensitivity of occupation matrix

[Stefano Fabris]

Dear Chan-Woo Lee,

if I understand correctly your mail, you are experiencing a common problem in DFT+U calculations. The issue is related to the existence of several local minima of the electronic solution present for the SAME crystal structure. One example of this for the case of FeO is provided in DFT+U example of the QE distribution (example 25 if I remember correctly). In that case, you can reach two electronic self-consistent solutions: one in which one electron is distributed over three d orbitals with almost equal occupancy, the other (with lower energy) in which the electron occupies only one d orbital. The calculation converges to one or the other depending on the initial values of the occupations. See the effect of starting_ns_eigenvalue

In my experience, the case of cerium oxides is even more delicate (and more frustrating the more you move towards complex crystal structures), but the way out is the same. Always carefully check the final occupancies, and in case of fractional occupations, play with starting_ns_eigenvalue and try to reach a lower energy solution with interger occupations on the localized orbitals entering the U term.

Hope this helps.

Stefano

On Feb 2, 2012, at 11:32 PM, Chan-Woo Lee wrote:

> Dear QE developers and users:
>  
> This is kind of follow-up question of my previous post: http://www.democritos.it/pipermail/pw_forum/2011-December/022823.html
>  
> I found that current Ce psp has some critical problems (for both CeO2 and Ce2O3) and tried to modify the psp. In doing so, I found Ce psp from Prof. Fabris (http://www.democritos.it/~fabris/Ce4_ps_f1d1-pbe-rc1.0.uspp.UPF ) and tried it with current O psp. The Ce psp is ultrasoft and the O psp is RRKJ. Both use PBE functionals.
>  
> In working with the psp, I found interesting results. I tried two cases: 1) SCF without any atomic relaxation 2) RELAX. I use QE 4.3.2 and Effective U is chosen as 5.0. Initial structures are adopted from http://prb.aps.org/abstract/PRB/v75/i4/e045121 . As relax from case 2) doesn’t do much, final structures from two cases are identical.
>  
> While the structures are identical, occupation of f orbitals are different and DOS are also not the same. For 1) case, occupation is almost 1.0 (~0.9) at single orbital state which is desirable, I think. Also, DOS shows that there is no partially filled f states near Fermi level which is the case for conventional DFT. So, Ce2O3 is predicted to be insulating. For 2) case, occupations are somewhat even: 0.23, 0.26, 0.29, and 0.12.. Also, electronic structure doesn’t change that much and there is still partially filled f states. Same atomic structures but totally difference electronic structures…
>  
> While I tried to understand and explain this to myself, I can’t find any good answers for this. I am currently testing “starting_ns_eigenvalue” and “mixing_fixed_ns” to see forcing certain occupations will be helpful or not.. but I am not sure.
>  
> Any comments or suggestions…???
>  
> -------
> Chan-Woo Lee, Ph.D.
> Postdoctoral Research Associate
>  
> Department of Chemistry
> University of Pennsylvania
> 231 South 34th Street
> Philadelphia, PA 19104-6323 
> 
> 
>  
> _______________________________________________
> Pw_forum mailing list
> Pw_forum at pwscf.org
> http://www.democritos.it/mailman/listinfo/pw_forum

---
Stefano Fabris
CNR-IOM DEMOCRITOS Simulation Center and 
SISSA - Scuola Internazionale Superiore di Studi Avanzati
tel: @SISSA +39 040 3787-872 @Elettra: +39 040 3758-735
email: fabris at sissa.it  
http://people.sissa.it/~fabris
---