LNONCOLLINEAR, ISPIN, MAGMOM, and SOC in MLFF

[reza_namakian1]

Dear VASP Team,

I have some questions regarding the use of LNONCOLLINEAR, ISPIN, MAGMOM, and SOC when calculating the energies of isolated atoms and dimers.
Currently, I am working on a perovskite material studied using VASP MLFF (https://doi.org/10.1103/PhysRevLett.122.225701).

1) I want to calculate the energies of isolated atoms as reference energies and include some dimer separations in the MLFF training. Should I use LNONCOLLINEAR, MAGMOM, and SOC for these calculations? Specifically, for Pb, I, N, C, and H, as well as for the dimers Pb-I, N-C, N-H, C-H, and I-H?

For example, in the MACE documentation, it is recommended to use ISPIN (https://mace-docs.readthedocs.io/en/lat ... ine-tuning):
“When computing your E0s, please use spin-polarized calculations.”

2) Do you have any recommendations for MAGMOM values for these elements and their dimers?

3) Given the high computational cost of these tags, I would prefer not to use them in the on-the-fly learning scheme for MLFF training on the bulk system, just for isolated atoms and dimers. Would this lead to significant inconsistencies? I’m not sure but based on a comment by Gábor Csányi here (https://github.com/ACEsuit/mace/discuss ... 5?sort=old), it might be fine. Could you please share your opinion with me on this?

Thanks!

Reza.

[martin.schlipf]

The only inputs that enter the training of the force field are the energy, the stress, and the forces. So as long as these are accurate, your force field should be as well.

Keep in mind that SOC is very important for the electronic structure so if you want to look at the band structure you need to use SOC.

For the molecules, the same considerations hold. Often you need to include magnetism if there is an unpaired electron. Typically the default choice the magnetic moment for a molecule is not so important because you have no magnetic site so the magnetization density will be distributed over multiple atoms.

[reza_namakian1]

Dear Martin,

Thanks a lot for your reply!

I'm mainly interested in the mechanical properties of the material. In that case, do you think I still need to include SOC for the material? As I mentioned in my previous post, I’m fine using it for dimers and isolated atoms, but I don’t think I have enough resources to handle a 100-atom supercell during on-the-fly learning with SOC or even ISPIN=2.

So, would it be a huge mistake to use different ab-initio settings for isolated atoms and dimers than the ones used for on-the-fly learning?

"Warning: It is very important not to change the ab-initio settings in the INCAR file between training from scratch and continuing training. Likewise, the POTCAR file should not be changed when resuming training."

Also, do you have any suggestions for MAGMOM values for the elements and dimers I need to calculate energy and forces for?

Many thanks for your advice on these!

Best,

Reza

[martin.schlipf]

I would not expect that SOC has a huge impact on the energies and forces but that is something that needs to be tested. You can simply run a training run without SOC and then pick some of the structures from the trajectory and recompute the energies and forces with SOC. If you observe that the error is small compared to the error that you make by using a force field, it should be fine.

I am not sure how much you would benefit from including the molecules and dimers if there local environment is quite different from the one you want to simulate later. If you just want a reference energy to compute formation energies, it does not matter if the setups are different as long as the energies are the same. The force field does not see your computational setup, it only sees the resulting energies, forces and stress. The advice not to change your setup stems from the fact that, in general, you can not assume that the energy will be the same if you change the setup. Then the force field would learn different energies than what you want to predict. If you convinced yourself that to different setups lead to the same energies, you can use the one that is computationally cheaper.

Regarding the MAGMOM values for the fragments: Again, these do not have magnetic sites, i.e., strongly localized magnetic moment. You expect that the magnetization density spreads out over the fragment. Therefore any MAGMOM setting should lead to the same result. The recommendation to set MAGMOM manually is mostly relevant if you have multiple magnetic sites and they may couple ferro- or antiferromagnetic. In these cases, VASP may not find the correct answer, if you do not start with the correct initial state. For your molecules, I expect no changes based on your starting magnetization as long as you do not set it to 0.