Stress Tensor comparison DFT/ML-FF

[julien_steffen]

Good morning,

we are conducting simulations of liquid metal surfaces using the VASP ML-FF. Currently, we are benchmarking the results of trained ML-FFs for several systems (e.g., liquid Ga, Sn, GaPt, etc.) by comparing them to direct AIMD simulations with VASP.
One observable, which is also of high interest for our experimental colleagues, is the surface tension of those systems. Therefore I want to compare surface tensions of DFT and ML-FF simulations for the same system at a range of different temperatures.
I tried to calculate them with the usual formula (with vacuum introduced in z-direction): r = 0.5 * Lz * (Pzz-0.5*(Pxx+Pyy)), where Pij are the respective components of the pressure tensor (where I assumed that I can use the stress tensor from VASP for this purpose).
I noticed, however, that the printouts concerning the stress tensor are quite different in the OUTCARs of AIMD and ML-FF simulations.
In the AIMD calculations (for a pure Ga surface slab), the OUTCAR section in each MD step is:

FORCE on cell =-STRESS in cart. coord. units (eV):
...
Total -20.47972 -35.62209 -24.26207 -4.37477 -3.16535 -0.29312
in kB -5.63510 -9.80159 -6.67583 -1.20374 -0.87096 -0.08065
external pressure = -7.37 kB Pullay stress = 0.00 kB

In the ML-FF calculations of the same system, the respective OUTCAR section is (I assume at least):

stress matrix after NEB project (eV)
16.76946 0.78240 4.79830
0.78240 28.29533 3.93928
4.79830 3.93928 19.15916

My question now is, how I can compare the values of both calculations directly, i.e., obtain pressure tensors from both models for further evaluation and calculation of the surface tension.
In the AIMD case, there is a `FORCE on cell =-STRESS` note at the beginning, so I assume that I need to take the values of the line `Total -20.47972 -35.62209 -24.26207 -4.37477 -3.16535 -0.29312` given in eV and then invert their sign?
In the ML-FF case, has the comment ` stress matrix after NEB project` any influence on the comparability, since no such `NEB project` is noted in the AIMD case?

Thank you in advance!

Best wishes,

Julien

[henrique_miranda]

Strange, I thought you would find the stress tensor written in the OUTCAR file both when running ML or AIMD.
Could you share your INCAR, POSCAR, POTCAR, KPOINTS, and OUTCAR files for the ML and AIMD runs?
(if you can show the issue on a system with few atoms then better)

[julien_steffen]

Ah sorry, you are right! I overlooked the respective "stress matrix after NEB project (eV)" section in the AIMD OUTCAR files.