DPLR: 2 different Wannier centroids ; bond types and fix dplr

[axgmz]

Hello!

Context and models trained:
I am trying to use DPLR for a slightly more complex system than the example bulk H2O by adding an anion. I have a pretty good dataset with coords, energies, forces and wannier centroids on O and my anion.
A short range DeePMD NNP with this training set is doing good. A DeepDipole model is also good at predicting the Wannier centroids.
I trained a DPLR NNP with a modifier. My Wannier centroid on O has a given charge, on the anion the number of Wannier centers averaged is not the same, the charge is different but this should not be a problem as the nuclei positive charges are also included in the modifier section of DeePMD input.

DPLR MD with Lammps:
I now have 3 real atom types: O, H and anion and 1 virtual atom (atom number 4 in Lammps data file) corresponding to the Wannier centroids on O and anion.

• I tried to have only one bond type for the WC-atom bond and 2 fix dplr in my Lammps input (see code snippet 1)
• I tried to have two different bond types for the WC-O and WC-anion bonds and 2 fix dplr in my Lammps input (see code snippet 2)
• I thought also that I could have only one bond type and 1 fix dplr but I did not manage to write a correct type_associate which should be something like type_associate (1,3) 4 bond_type 1 but I cannot guess how to write it correctly

=> Nothing works, the two first tries output the same error: ERROR: find a bonded pair the types of which are not associated (/home/conda/feedstock_root/build_artifacts/libdeepmd_1660884873849/work/source/lmp/fix_dplr.cpp:208)

Code snippet 1:

fix             wcO all dplr model dplr.pb type_associate 1 4 bond_type 1 # for O - WC
fix_modify      wcO virial yes
fix             wcanion all dplr model dplr.pb type_associate 3 4 bond_type 1 # for anion - WC
fix_modify      wcanion virial yes

Code snippet 2:

fix             wcO all dplr model dplr.pb type_associate 1 4 bond_type 1 # for O - WC
fix_modify      wcO virial yes
fix             wcanion all dplr model dplr.pb type_associate 3 4 bond_type 2 # for anion - WC
fix_modify      wcanion virial yes

Do you have any experience with different atoms associated with their own Wannier centroid? From the fact that model_charge_map is a list in the DPLR model training json file it seems like it should be somehow possible. Can I have two fix dplr in my Lammps input? Should I have two bond types? Only one?

Thank you for your help!

[ChunyiZh]

Hi,

I would suggest that you set the Wannier centroids (WCs) for different atom types also as different types. For example, if your system is NaCl solution. You can set O as atom type 1, H as atom type 2, Na as atom type 3, Cl as atom type 4, the WCs of O as type 5, the WCs of Na as type 6, and the WCs of Cl as type 7. Then, there are 3 bond types, bond type 1 between atom types 1 and 5, bond type 2 between atom types 3 and 6, and bond type 3 between atom types 4 and 7. In Lammps input, just one "fix dplr" is enough.

Following are some settings for this system.

Training DW model:
"fitting_net": {
"type": "dipole",
"sel_type": [
0,
2,
3
],

Training DPLR model:
"modifier": {
"type": "dipole_charge",
"model_name": "dw.pb",
"model_charge_map": [-8,-8,-8],
"sys_charge_map": [6, 1, 9, 7],
"ewald_h": 1.0,
"ewald_beta": 0.40
},

Lammps input:
neigh_modify every 10 delay 0 check no exclude type 1 5 exclude type 3 6 exclude type 4 7
group real_atom type 1 2 3 4
group virtual_atom type 5 6 7
fix 0 all dplr model model_lr.pb type_associate 1 5 3 6 4 7 bond_type 1 2 3

[axgmz]

Thank you very much for this clear and detailed answer! It helped me a lot and solved my problem!