Inconsistent energies in 6.4 MLFF for different supercells

[johkl_]

Dear developers,

I've noticed some inconsistencies in the energies predicted by the new vasp 6.4 MLFF (ML_MODE = refit). The predicted energy per atom differs substantially between different supercells representing the same structure. This does not seem to happen for "simple" expansions (eg. nx x ny x nz) of the supercells for which the structures where fitted, but only for more "complicated" supercells. For the older ML_MODE = refitbayesian, the energy per atom seem to be consistent for all possible supercell expansions I've tried.

I've reproduced this error in at least 2 different systems, I attach an example based on the "Liquid Si" example from the wiki; Taking the input files from the wiki and using the ML_AB file obtained at the end of the run to produce two ML_FF files (ML_MODE = refit and ML_MODE = refitbayesian) I then evaluate the energy on the 64 atom starting supercell from the wiki, and a supercell with the header:

non-diagonal supercell
10.861800000000
1.00000000000000 3.00000000000000 4.00000000000000
5.00000000000000 1.00000000000000 3.00000000000000
1.00000000000000 3.00000000000000 2.00000000000000
Si
1792
Direct coordinates

For the ML_MODE = refitbayesian case the energies of these two supercells are consistent:
-.33846532E+03/64 =-5.288520625 eV/atom
-.94770289E+04/1792 = -5.288520591517857 eV/atom
while for the ML_MODE = refit case they are not:
-.33871124E+03/64 = -5.292363125 eV/atom
-.91807776E+04/1792 = -5.1232017857142855 eV/atom

Input files to reproduce this error are attached, along with certain output files that could be fit into the forum size limit.

Best,
Johan

mlff_bug.zip

[ferenc_karsai]

Thanks for the detailed report.

One thing could you please upload the ML_AB file that you used? I will create my own but it's good to check with your data too.

[johkl_]

Please find the ML_AB file in the attachment.

mlab.zip

[andreas.singraber]

Dear Johan,

thanks again for putting a lot of work in your very well-prepared bug report! It helped us to quickly identify that there is a problem with the cell list algorithm when "very" triclinic lattices are used (angles much larger or smaller than 90 degrees). The erroneous MLFF predictions only appear with ML_MODE = run and a force field suitable for the fast-execution mode (refitted with ML_MODE = refit). A corrected cell list algorithm will be included in the next patch release of VASP. We are very sorry for the inconvenience! In the meantime we prepared two workarounds which are documented on our "known issues" Wiki page:

Known issues date stamp: 2023-05-17

Option (1) is highly preferred because the other one will slow down the MLFF predictions significantly. I hope this solves your problem for the moment, please let us know if there are further issues!

Thank you!

All the best,
Andreas Singraber