Question about "Whether Surface Slab thickness need to be larger than Rcut"

[roger13231]

Dear Developer and colleague,

I have a question about Rcut and surface model thickness selection during surface potential training.

A reviewer had an opinion on an article that my senior participated in that, physically speaking, if the thickness of the interface slab is less than Rcut, then the interaction of atoms at the top surface will be affected by the vacuum above and a small portion of the vacuum layer below. This "small portion of the vacuum layer below" could cause a slight different influence on predicting the interaction of those top surface atoms.

I am wondering what this suggestion sould like or whether I could continue use my current training data, which slab is smaller than Rcut. I don't physcially understand this influence and what the difference on considering this small portion of vaccum or not.
If the thickness needs to be increased, it seems like it would require many layers, possibly up to 7A. I see some current colleague in DP discussion group mention they use 4 or 5 metal layer, which is maybe 4-6 A.

[njzjz-bot]

The slab thickness does not have a strict requirement to be larger than rcut.

For a local DeePMD descriptor, an atom uses neighbors that actually exist within rcut. A surface atom correctly sees vacuum on the exposed side. Making the slab thinner than rcut does not introduce “vacuum neighbors”; it only means that the atom may see all layers of that finite slab.

There are two separate convergence conditions:

1. The vacuum separation between periodic slab images should be larger than rcut (preferably with a margin), so atoms do not acquire unintended neighbors across the periodic boundary.
2. The slab thickness should be converged for the physical observable: surface energy, adsorption energy, forces, relaxation, electronic structure, etc. It must be thick enough that the two surfaces or the constrained bottom layers do not materially affect the surface of interest.

Therefore, a 4–6 Å slab is not automatically invalid merely because rcut is 7 Å, and increasing it to exceed rcut is not by itself a proof of convergence. Compare the target quantities and surface forces for successively thicker slabs, and include the relevant slab environments in training and held-out testing. The periodic-image distance, not the slab thickness alone, is the cutoff-related geometric check.

Coding agent: Codex
Codex version: codex-cli 0.149.0
Model: gpt-5.6-sol
Reasoning effort: xhigh