[Pw_forum] strange behaviour in LDA+U calculations

[Matteo Cococcioni]

Dear Giuseppe,

first of all: LDA+U is variational and the energy should be minimized 
during the
electronic convergence independently from the algorithm you use.

the answer/solution to your problem, however, is probably not easy. Of 
course it depends
on many details (which are not in your email) about the oxidation state 
of Co, the splitting
between d orbitals, the magnetic state.
To me, from what you describe, it looks like it could be a case where a 
multi-reference wavefunction is
needed to describe the ground state as it happens to be the case (I 
believe) for some TMOs when
they become paramegnetic still preserving their Mott insulating 
character (LDA or GGA would predict a metal)
This paper mentions this problem for some of the high pressure of MnO
arXiv:1006.0565v1 [cond-mat.str-el] 3 Jun 2010
LDA+U works when you have a single determinant or when you break the 
symmetry to precipitate the system
in one of those solutions. This is good when these solutions are 
degenerate, a bit worse when they are not....

so, at first I would try to understand what is the degeneracy situation 
in your system and how the different
solutions between which your system oscillates differ from each other in 
terms of occupation of d and p orbitals.

hope this helps.

Matteo



Giuseppe Mattioli wrote:
> Dear all
> I'm facing strange problems while attempting to perform LDA+U calculations on 
> Co-O clusters. PBE calculations work fine, both with PW and CP, both in 
> closed shell and LSDA framework. In particular, ferromagnetic or 
> antiferromagnetic initial guesses lead to closed shell electronic structures, 
> i.e., total and absolute magnetization equal (or very close) to zero, and a 
> safe insulating system.
> In order to improve the p-d hybridization between O 2p and Co 3d orbitals 
> (their contributions to the DOS are quite separated in PBE calculations, and 
> they are expected to be strongly mixed near to the valence band maximum)
> I've performed some PBE+U calculations by using reasonable values for the Co 
> 3d and O 2p channels. In the case of closed shell calculations I obtain quite 
> reasonable results, i.e., equilibrium geometries nicely similar to PBE ones, 
> and an improved DOS in a better agreement with the available experimental 
> data. I've tried to perform open shell calculations since January, I've tried 
> to apply all of the standard and less standard tricks, but I have never (!) 
> been able to converge a single calculation. There are no errors, they simply 
> does not converge, with the system magnetization dancing around some selected 
> values, depending on the initial guess (fm or afm). Notice that the system is 
> still insulating (at a PBE+U closed shell level) and I've tried everything 
> from fixed occupations to quite high smearing (gaussian, cold, ...) values.
>
> I thought that the CP code could help me, but I'm facing an even more strange 
> problem... The very robust CP minimization of electronic dof successfully 
> reach some ground state at the PBE+U level, the only (minor) problem being 
> that you must put into the inpfile the final magnetization of the system. 
> Then, I've tried to perform both damped and NVT (with Nose thermostats) MD 
> simulations, trying several different open shell (nspin=2) configurations, 
> ranging from tot_magnetization=0 to tot_magnetization=20. In the case of 
> damped MD, the system temperature was kept very low, but my Co-O clusters 
> tried to split into pieces (they are quite stable in both PBE and PBE+U PW 
> calculations) while the total energy lowers. In the case of NVT MD 
> thermostats went bananas and my clusters melt. You can scrub away Co and O 
> scattered atoms from the supercell walls...:-)
> In the case of PBE calculations (both damped and NVT MD) everything went fine 
> (no explosions, reasonably fast thermalization of the investigated systems, 
> averaged bond distances similar to PW ones...) 
> I can provide input files, but I suspect that this is not a matter of input 
> failures... Is it LDA+U variational? I suspect it isn't... Is it foolish to 
> perform MD calculation in such framework? Could you lend me a heavy hammer to 
> persuade PW open shell calculations into convergence...?
>
> Thank you in advance for suggestions
>
> Giuseppe
>
>