Structure relaxes to very different magnetic states on first and second relation run

[sophie_weber]

Hello,

I am trying to roughly simulate a (001) slab of antiferromagnetic chromia with a Pt layer deposited on top, and I'm trying to relax the internal positions (leaving the in-plane lattice vectors fixed to relaxed Cr2O3) with collinear magnetism. After the first relaxation, when I look at the projected magnetization in the OUTCAR (the full file is too large to attached so OUTCAR_relax1 just has the final part after ionic minimisation is reached) it has converged to what I believe is the correct magnetic ground state (chromia retains its AFM 'up down up down.." order along the [001] axis, and the layer of 3 Pt on top acquires a very small (<<1 uB) magnetic moment.

The weird part happens when I copy the CONTCAR to a new folder and start a 2nd relaxation (relaxrun_2.tar) using the exact same input files with the relaxed structure (I usually do this as a sanity check, and if it's relaxed properly, it should finish after one ionic step). However, the relaxation runs for around 50 ionic steps, and when the force criterion is again reached, the new relaxed structure (called CONTCAR_2ndrelax ) very similar to the input relaxation structure, however, now according to the new OUTCAR (OUTCAR_relax2) the site-projected magnetization has converged to something completely different and nonsensical, with many Cr now ferromagnetic coupled and some becoming almost nonmagnetic, even though they are at the other side of the slab and thus should not interact significantly with Pt. (I have done very similar relaxations with chromia using this same thickness of slab and e.g. using a monolayer of Fe rather than Pt, and nothing like this has ever happened, and the layer-projected DOS looks bulk-like in the center of the slab, so I don't believe it's a thickness issue).

As a side note, I am aware that VASP 5.4.4 often does not converge to the correct magnetic ground state due to a symmetrisation bug, however, since I have symmetries turned off (isym=0) this should not be an issue. And furthermore, even if it was the problem, it makes no sense that the structure would converge to the correct ground state on the first relaxation and not on the second.

I've attached tar files with the inputs and outcars for the two calculations; any help or insight would be greatly appreciated. Thank you!

[alexey.tal]

Dear Sophie,

Do you start the second relaxation run from the converged charge density? If you start the second relaxation calculation from the atomic charge density, it might be that the calculation couldn't converge to the same magnetic configuration you found in the first calculation.

I would suggest that you try running your second relaxation starting with the converged charge density and wave functions (CHGCAR and WAVECAR) of the first relaxation calculation. Also, setting the density mixing AMIX to a smaller value should help the convergence.

[sophie_weber]

Thank you! I will try this.