Pretrained Si NEP Models

[tomasfbouvier]

We are working with silicon (Si) containing impurities. We already have a NEP model that performs well for pure Si. Would it make sense to freeze this pretrained Si model and then train additional parameters specifically for the impurity-related interactions, such as Si-X and X-X (where X represents the impurity)?

[brucefan1983]

This is not a very good idea I think. When you have extra elements, the existing parameters do not necessarily the optimal ones any more. For example, the nueral network parameters should adapt from pure silicon environments to also account for mixed environments. The best you can do is to keep that old NEP model for silicon and develop a new one for SiX, but reusing the Si dataset as part of the expanded dataset.

[tomasfbouvier]

Hi, thank you for your answer. I would believe that the parameters for Si are completely different in the case of a homogeneous stochiometry 1:1 but when having single atoms within a matrix I would not expect the parameters for the matrix to be that different. I think another user had already suggested the possibility of augmenting a "pretrained" model to avoid starting from scratch.

If I understood correctly you suggested overlaying two potentials: one for Si and one for SiX (the latter one being trained over the residual of the energies forces and virials computed by dft and the former potential). I am trying that currently and the results seem ok. Is it possible though to avoid overlaying and assign one pair to each potential? Si-Si for nep1 and Si-X and X-X to nep2?

[brucefan1983]

I see. There is way to do what you want, but it requires changing theh source code a little. Do you really want to delve into it?

[brucefan1983]

It requires to:

• change the source code such that the descriptor normalization will be read in from nep.txt.
• prepare a nep.restart file such that the Si-related trainable parameters will be frozon (simply put the variance in the second column to zero for those parameters)