Training NEP with systems having data from unit cell and replicated cell data

[Achyut-blip]

Hi everyone,

I have been using NEP + GPUMD for my research purpose.

I was wondering if we can train the NEP model by merging the aimd data for a unit cell structure and replicated structure. How does the train.xyz file look like in this case? Can we specify the multiple train.xyz file while training? Or we can combine the datasets for different structures?

I would love to get response in this regards
Thank you for your attention to this matter.

[brucefan1983]

The train.xyz file can contain multiple stuctures with the same or different sizes. If you have two datasets such as train1.xyz and train2.xyz, you can combine them with

cat train1.xyz train2.xyz > train.xyz

[Achyut-blip]

Thank you so much. It helps alot to me.

One more question, if I am to train on different atom types in two different train.xyz files,
how does the nep.in file look like?

In the nep.in file

Can I simply write

type 5 a b c d e
For a, b , c, d, e elements that are in total of my train.xyz file?

Does the order of elements in different train.xyz file matter while specifying this section of nep.in file?

Thank you for your attention to this matter.

[brucefan1983]

You can specify as you wrote, and the order can be arbitrary. The potential file nep.txt will record the order you chose.