When I was encountering convergence problems with magnetic systems the attached PPT by G. Kresse was VERY helpful. Particularly slides 20-22

Start by looking at the standard out or OSZICAR file to determine the type of convergence failure and in the OUTCAR file for the GAMMA value. I suspect changing AMIX and BMIX will help.

Dear,

Did you try to consider some evolutionary based algorithm, e.g. Differential Evolution, Particle Swarm Optimization?

Are these implemented in VASP?

Hi Martin,

Fot a quick test, you may want to edit your MAGMOM line a bit. You say the eta phase is not ferromagentic, but it does consist mostly of Ni. While there are no parallel aligned Ni moments in the eta phase, the Ni moments may not be suppressed entirely (in the way fcc Fe is not ferromagnetic, but the Fe atoms do definitely have a local magnetic moment)? So 64*0 and 64*5 may both be very far from what the system would want to converge to. You may want to try a MAGMOM line in which some atoms in the eta part have spin up, others spin down.

Also, a moment of 5 Bohr magneton is pretty large. I know that the VASP manual recommends large initial moments. But I think that was done because if a calculation converges, a large initial moment will usually be reduced, while a small initial moment won't always grow larger. So a large initial moment is 'safe'. But you need to get to convergence first. In a 4x4x4 fcc cells Ni calculation I did, the supercell had a total moment of 154 BM, so less than 1 BM per atom (calculated within spheres around the nuclei with the default radius in the POTCAR). 5 BM as initial moment may be starting off the system in a rather unnatural magnetic state.

When I was encountering convergence problems with magnetic systems the attached PPT by G. Kresse was VERY helpful. Particularly slides 20-22

Start by looking at the standard out or OSZICAR file to determine the type of convergence failure and in the OUTCAR file for the GAMMA value. I suspect changing AMIX and BMIX will help.

@Erik

Thanks a lot, I'll try to play a bit more with AMIX and BMIX.

Actually, by now I did find a combination of parameters that converges, but very slowly, usually needing about 180 DAV iterations

@Peter

Thanks a lot, reducing the MAGMOM to 72*3 64*0 helped

@ Erik NyKwest

Thanks a lot for your reply to Martin's question, which also helped me!

But I am not quite understanding some points of your answer since I have the same problem to get the electronic converge for Ni MOF calculation. (I have 6 dimer-Ni, so my input of MAGMOM=6*2 6*-2)

"Start by looking at the standard out or OSZICAR file to determine the type of convergence failure and in the OUTCAR file for the GAMMA value. I suspect changing AMIX and BMIX will help. '

(1) How to judge the convergence failure by mag value in OSZICAR file 

(2) "AMIX and  BMIX" will help, how about AMIX_MAG and BMIX_MAG tag? Do you understand the physical meaning of these parameters since I can not find a resource to find the solution?

I am highly appreciated your any reply and help. Since the magnetic converge problem has taken me almost more than one month to figure it out.

@Wenqin You

I'm glad it helped! To clarify, when I say to look at the STDOUT or OSZICAR I mean look at the total energy and energy change as a function of the electronic step number (N). As N -> infinity the change in energy should go to zero and the total energy should converge to a constant value. If the change in energy starts GROWING or diverging your charge mixing is too strong. In the case of linear mixing this means reduce AMIX. If it is converging slowly, sometimes increasing the mixing will reduce the calculation time. The mixing is optimum if the average value of GAMMA is 1. Watching the value of gamma can also give you clues if something is going wrong.

As for AMIX_MAG and BMIX_MAG, I'm sorry to say I've never used them. I will suggest however that you check you MAGMOM setting and make sure you are applying the spins to the correct atoms. As you have it set the first 6 atoms in your POSCAR are spin up while the next six are spin down.

@ Erik Nykwest

Thanks so much for your reply. It helped me a lot. 

Sorry if this is off-topic, but we're collecting systems which are hard to converge in DFT to form a test suite (the SCF-Xn test suite) to benchmark new and existing algorithms for finding the Kohn-Sham ground state. If you'd like to try it out and/or contribute, you can find it on GitHub here:

https://github.com/NickWoods1/scf-xn-testsuite

and the results of trying it for a range of current algorithms have recently been published here:

Article Computing the self-consistent field in Kohn-Sham density fun...