How to exclude specific interactions at interface for SHC calculations?

[pedramonline87-ship-it]

Dear DeepMD developers,
I am working on calculating the Spectral Heat Current (SHC) at interfaces using DeepMD potentials, and I have a question regarding interaction handling at boundaries.
For accurate SHC calculations at interfaces, it is necessary to exclude certain interactions across the interface boundary. For example, when calculating the SHC of silicon/diamond structure at an interface, the Si-Si interactions across the interface boundary should not be included in the heat current calculations, while interactions within each domain should be considered normally.
My questions are:

Does DeepMD-kit have a built-in feature to selectively exclude or mask specific atomic interactions at interfaces or boundaries during MD simulations?
If such functionality exists, could you please point me to the relevant documentation or provide an example of how to implement this?
If this feature is not currently available, would it be possible to implement this through custom modifications, or are there any workarounds you would recommend?

I understand that DeepMD focuses on machine learning potentials, but controlling interaction inclusion/exclusion seems crucial for interface thermal transport calculations.
Thank you for your time and for developing such a powerful tool!
Best regards

[dosubot[bot]]

Hi @pedramonline87-ship-it! I'm Dosu and I’m helping the deepmd-kit team.

DeepMD-kit does not have a built-in feature to exclude or mask specific atomic interactions at interfaces or boundaries during MD simulations. The exclude_types option can be used during training to remove contributions from specified atomic pairs, but this exclusion is global and not spatially selective—it cannot be applied only at an interface or dynamically during a simulation. The many-body nature of DeepMD potentials makes runtime pairwise exclusion technically infeasible, and attempts to use LAMMPS commands for this purpose are ignored by the DeepMD model (discussion) (user experience) (technical explanation).

Workarounds include training separate models with and without certain interactions and using free energy perturbation or post-processing to analyze the effect, but these approaches require custom workflows and are not built-in features (discussion). For SHC calculations at interfaces, you may need to post-process the heat current data to exclude contributions from undesired interactions, or design your training data and model exclusions carefully to approximate the desired effect.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}