Relaxation of magnetic heterostructure with SOC

[william_schaarman]

I'm relaxing a Cr2O3-Pt(111) slab Heterostructure. Both materials have been relaxed beforehand with Pt strained to fit the lattice parameters of Cr2O3. During the relaxation process of the heterostructure I have started with different distances between Cr2O3 and Pt. I'm aware of a minimum that was found by previous calculations around the 2.05Å (z-distance between the top Cr atom and bottom Pt atom), hence I have chosen to start with a distance of 1.7Å, 1.8Å, ... , 2.3Å, 2.4Å between the two materials. Consistently across all relaxations in the first ionic step the gap between the structures increases drastically. After this first step the relaxation slows down over the next few steps into a local minimum. An overview of the relaxation can be found in the following figure:

RelaxCr2O3Pt.png

My question about this is twofold. First I wonder what what can be the main cause behind the large jump in the first step. Secondly I would appreciate advice on how to converge to the global minimum more consistently. From what I checked 2.05Å is closer to the global minimum than 2.18Å, which is a common local minimum. I have already tried IBRION=1,2 which makes minimal difference from my quick test.

In the zip added I've placed both the POSCAR for the converged 2.05Å case and the start for the 1.8Å case. The output file by vasp only goes over the second batch of ionic steps as the cluster I work on requires a resubmission of jobs after 24h of computing.

[henrique_miranda]

Hi,

1. The reason you get a large jump in the first step is because for your calculation settings (xc-functional, ENCUT, POTCAR, KPOINTS) the initial separation you are using is insuficient. The layers feel a large force and the ionic minimizer tries to pull the layers appart before eventually relaxing.
2. From this image alone I am unsure what the issue is. Do the different calculations converge to different layer separations depending on the initial separation you use? It looks like but since the calculations stop I am not sure. Can you try plotting the total energy for a single shot calculation (no ionic minimization) as a function of the layer separation? Perhaps the energy landscape is not deep enough to actually bind the layer so it is just being pushed away and then never attracted again. You might need to include some type of van der Waals interaction (but I am not familiar with this system you are studying).

A few tips on performance for faster testing:

1. you can probably do an initial relaxation with lower ENCUT
2. ADDGRID=.TRUE. adds some computational time, it might reduce the noise in the forces but for an initial test is maybe overkill
3. NCORE=32 is probably too large, I would suggest a value of 8 or so (Before running the ionic relaxation you can actually time a single SCF step by setting NELM=1 and IBRION=-1 and check what is the optimal value of NCORE to use for your particular setup).