Incompatible between "kspace_style pppm" of LAMMPS and NequIP

[gly123456-phy]

Hello,
I would like to ask an error when I run LAMMPS on a trained model. After adding the line "kspace_style pppm" in the input of LAMMPS, there is a reported error:
"PPPM initialization ...
ERROR: Kspace style requires atom attribute q (src/KSPACE/pppm.cpp:201)
"
In LAMMPS, "kspace_style pppm" is used for long-range Coulombic interactions and requires per-atom charges. Is it possible that I can also assign atomic charges using NequIP-trained model?

[MarcD3]

Hi,

NequIP does not natively have any support for charges that would work with pppm (normally you would just use pair_style nequip, and no kspace). If you were to overlay charge interactions by assigning the LAMMPS q value in some way in addition to the NequIP model predictions, you would need to ensure that this physically made sense, as NequIP has already learned the complete potential energy surface to the best of its ability while training to DFT energies and forces.

--Marc

[gly123456-phy]

Hi @MarcD3,
Thanks a lot for the clarification. Can you let me know the format of overlaying charge interactions on top of NequIP model? I guess this step is done in the input of LAMMPS?

[MarcD3]

Note that adding any charge interaction in addition to a standard NequIP model is not the standard use, so unless you are developing a new method, it should not be done. But, adding any sort of charge interaction would have to be handled externally to the NequIP model. For example, if you use the "charge" atom_style in LAMMPS, you could assign charges to each atom in a lammps data file, and read that file in during the lammps script. Then, the kspace pppm would be able to compute some interactions based on those assigned charges. Though, if you add charge during inference time like this, and a consistent type of charge interaction was not present when you were training the NequIP model, I don't expect this would produce an accurate simulation, as otherwise this would effectively be double-counting the electrostatic interactions to some degree.

--Marc

[gly123456-phy]

@MarcD3 Got you! Thanks a lot for your comment!