ERROR: pair_mace does not support vflag_atom when run lammps-mace

[liubaoshuai1402]

Dear developers,

I want to calculate the thermal conductivity of my bulk phase by MACE potential using the Green-Kubo (GK) formula in LAMMPS.

But lammps report an error: pair_mace does not support vflag_atom when i use cpu-calculation.
lammps report an error: mace/kokkos eflag_atom and/or vflag_atom not implemented. (src/KOKKOS/pair_mace_kokkos.cpp:82 when i use gpu-calculation.

I am not a lammps expert, but if vflag_atom and eflag_atom are frequently used in lammps, why mace don't support them in lammps. Is it a technical problem? Do you plan to add this support in the future?

Is there any alternatives to calculate the thermal conductivity by MD methods using MACE potential?
If you need my input file of lammps, i will supply.

vflag_atom will be set in lammps when calculating per-atom stress which is need by the Green-Kubo (GK) formula

[ilyes319]

I guess vflag_atom means atomic virials? This not computed right now in MACE. I think there is a PR here: you could give it a try:#528

[liubaoshuai1402]

I guess vflag_atom means atomic virials? This not computed right now in MACE. I think there is a PR here: you could give it a try:#528

I have tried the custom MACE, but one thing I want to consult with you.
Since Iammps uses MACE potential to calculate atomic virials, should I consider virials in my training set, or just energy and forces is enough?
If virials are needed and the weight of forces an energy is respectively 10 and 1, how much the weight of virial should be set?

[gabor1]

If you have virials (stresses) in your training set, it's always good to include them

[MicPellegrino]

Hi, I am encountering a similar error when trying to compute stress/centroids per-atoms in LAMMPS.
I initially thought it may have something to do with the newton on option, but it doesn't seem to be the case (it won't be specific for MACE otherwise).
I might be a newbie, but I don't really see how #528 could solve this in any way. Shouldn't something be fixed in the MACE-LAMMPS (or rather MACE-Kokkos) C++ interface instead?
Also, for some reason compute centro/atom works on CPU but not on GPU (compute stress/atom doesn't work on either).

[ilyes319]

Now you can use the following interface: https://mace-docs.readthedocs.io/en/latest/guide/lammps_mliap.html
It include atomic virials.