ZBL parameters and VASP PPs; Active learning and passivation

[TypicalBarmalei]

1. Advice on Choosing sw_rmin and sw_rmax for ZBL

I’m using the se_e2_a descriptor to model high-energy ion/atom impacts, where small interatomic distances between atoms may appear. The DeePMD-kit v2: A software package for Deep Potential models paper mentions that ZBL potentials can be combined with DeepMD to handle such cases.

Proposed Approach:

    sw_rmax = distance, above which reproducibility of VASP calculations is more or less guaranteed.

    sw_rmin = distance, below which we're highly unsure that VASP results are reliable.

Is this reasoning sound? While DeePMD uses softmin distances for boundaries of interpolation (effectively pulling the value closer to less distant neighbors), it still would be nice to associate it with VASP PAW’s "safe" distance (like sum or PP RCOREs of interacting atoms). Perhaps, there are better ideas?
Also, as I've read in other discussion, per-interaction-type definition for ZBL parameters is not implemented (yet?). But, if we have a "big" atom (like Fe, Ti...) and a "small" atom (like O, N, C...), their RCOREs will differ drastically. And safe distances for O-O bonds will probably be very dangerous for the metal bonds.

2. Is Passivation of Dangling Bonds Necessary for Force/Energy Accuracy? Or atom_pref.npy can be used.

DP-GEN uses max. force deviation to sample structures for DFT labeling, and same structures as in MD are then used in DFT. But some AL papers focus on extracting the atomic environments of "bad" atoms instead and surrounding them with a vacuum.

Question:
For systems with dangling bonds (metals, semiconductors), is passivation (with Hydrogen, perhaps) required to ensure reliable force/energy predictions from DFT? Does this depend on the used descriptor type?

Or, as I've seen, DeePMD has the atom_pref.npy feature, allowing me to set 0 weights on the "outer" atoms. But would energy need a correction in this case? I've also seen the atom_ener related parameters, but couldn't figure out its' usage in my case.

3. Best Descriptor for Deposition/Irradiation Modeling?

Which descriptor is more suitable for such nonequilibrium processes? Model compressibility would be a great advantage, of course.
Context: Modeling high-energy processes (e.g., sputtering, implantation).

Would love to hear others’ experiences!