Is there a minimum size of atoms needed in the cell

[dominicvarghese]

Is there a minimum size of atoms needed in the cutoff radius while doing MACE calculation with the foundational models? When I was looking at the dipole calculations with strain on a R3m BaTiO3, I was getting a different value for the unit cell while getting the same value for the different sizes of supercell - 2x2x2, 3x3x3, 4x4x4, 5x5x5 .

The unitcell POSCAR is as follows:

Ba Ti O 
 1.0000000000000000
     2.8173090586611149   -4.8797226184797458   -0.0000004024366094
     2.8173091662559111    4.8797226805995750    0.0000003250012582
    -0.0000000953031830    0.0000005166930899    6.9144485699551312
 Ba  Ti  O  
   3   3   9
Cartesian
  2.8173091861102568  1.6265743393275138  4.6005776018218807
 -0.0000001600243227  0.0000007035559735 -0.0090550861919234
  2.8173090693235343 -1.6265743758181954  2.2957614229222205
  2.8173095424503138  1.6265744252572643  1.2075076872235959
 -0.0000000063313679 -0.0000000120364665  3.5123236635227264
  2.8173092001215165 -1.6265724060001847  5.8171405519122850
  2.8173088968920386  3.2362411643605058  2.2228611287714233
  1.4232964327976985  0.8217404599396383  2.2228609330500415
  4.2113211585025736  0.8217406541439330  2.2228608793506726
  2.8173088818856504 -3.2700554442939915  4.5276763344698745
  4.2406055640879750 -0.8048331014593364  4.5276771676414667
  1.3940125503393066 -0.8048334000009266  4.5276773389224578
  2.8173090742475368 -0.0169061020807874  6.8324939168724343
  1.4232962712027313 -2.4314071397295827  6.8324929600290183
  4.2113219194454725 -2.4314072801233841  6.8324929339847458

So here in the unitcell within the cutoff radius, some atoms from the periodic images would be coming in, could this be the reason for the error in this case? Or could it be something else?

[gabor1]

MACE takes periodic boundary conditions into account, regardless of the size of the unit cell, there is no restriction of any kind.

I don't quite understand your question. What are you exactly calculating with the mace foundation models? not the dipole moment presumably.

[gabor1]

If you are relaxing the system, then maybe your geometries are different in the unit cell and supercells because there is an unstable phonon mode?

[dominicvarghese]

I was trying to get the piezoelectric coefficient by applying strain and relaxing the cell. MACE was used to do the atomic relaxations in the individual cells. To get the polarisation, I used BEC from doing a VASP run for the unitcell for calculating the polarisation.

I also checked the irrep modes in both unitcell and the supercell and it was same in both cases using ISOTROPY suite. Both cells were still R3m space group.

Bulk | VASP DFPT (Ionic only) |  3.779
1x1x1 | MACE-R2SCAN | 3.910
2x2x2 | MACE-R2SCAN | 3.730
3x3x3 | MACE-R2SCAN | 3.730
4x4x4 | MACE-R2SCAN | 3.733
5x5x5 | MACE-R2SCAN | 3.730

The last column is the zzz piezoelectric coefficient.

[gabor1]

Ok, so let's split the problem in two. Forget the vasp calculation, let's focus on why you get different geometries in the unit cell and the supercell after the mace relaxation. Are you using float64? Calculate the distance between the corresponding atoms to see what the actual deviation is. Do a phonon calculation in the unit cell to check for negative modes.

[dominicvarghese]

After running through the diagnostics, it turns out the discrepancy wasn't coming from the MACE relaxations or the physics of the cell at all. The relaxed geometries of the $1\times1\times1$ unit cell and the $2\times2\times2$ supercell are actually identical.

The error was entirely on my end in my post-processing script. When I was calculating the atomic displacements between the unstrained and strained cells to compute the change in polarization, I wasn't properly mapping the atoms. The script was comparing atoms from completely different positions (index jumbling), which threw off the calculations when multiplied by the Born Effective Charges.

I re-ran the relaxations explicitly enforcing float64. The resulting structures were identical to the previous runs (differences only appeared at the 4th decimal place), ruling out any precision artifacts.

Now the piezoelectric coefficients are well converged with the supercells,

Bulk | VASP DFPT (Ionic only) |  3.779
1x1x1 | MACE-R2SCAN | 3.913
2x2x2 | MACE-R2SCAN | 3.913
3x3x3 | MACE-R2SCAN | 3.913
4x4x4 | MACE-R2SCAN | 3.913
5x5x5 | MACE-R2SCAN | 3.913

I used the MATPES-R2SCAN model for all the MACE relaxations, whereas the original VASP baseline used PBEsol.

I also made the phonon dispersion, with Phonons(atoms, calculator, supercell=(4,4,4), delta=0.01)

Thanks for the suggestions which helped with catching this error.

[gabor1]

You're welcome! (But there appear to be unstable modes!)