Fine-tuning on multiple data sets

[hyeon4977]

Hello, I'm trying to fine-tune an OpenLam-based .pt file.
I'm working on three aimd datasets.

There are five types of atoms in total. Dataset 1 uses only four atoms, while the other two use all five.

In this case, when creating the type_map and type for dataset 1, should I unify the representation format with the other two data sets, such as 0, 1, 2, and 4? Or should I recognize the atoms in each folder separately? Or is padding with ghost atoms using -1 the correct way to do it?

example) Data 2 and 3 use C H O N Si, but if data 3 only contains C H O Si, is it correct to use 0, 1, 2, 4 for type.raw?

I look forward to your response.
Thank you.

[iProzd]

Hi @hyeon4977, that's correct — as long as each dataset is internally consistent, meaning type.raw matches its own type_map.raw (e.g. type_map.raw = C H O Si with type.raw containing 0,1,2,3, or type_map.raw = C H O N Si with type.raw containing 0,1,2,4, see here for detailed format), and you just need to ensure the type_map in input.json covers all element types in your data so the mapping works automatically. One more thing, for automatical way to convert data format (e.g. from vasp/lammps/cp2k etc.) for DeePMD-kit, see dpdata. Hope this helps.

[hyeon4977]

Unfortunately, I can't convert using dpdata.

1. input.json contains all the elements of the dataset.
2. Suppose I'm using three datasets. The first database maps Si, N, and O to 0, 1, and the second database maps N, O, and S to 0, 1, and 2. Is it correct to simply fill in the type.raw and type_map.raw files within each database?
3. When manually converting OUTCAR to dpdata, the atomic coordinates and cell box sizes are slightly different. Why is this?

Sorry for the long follow-up question :)

[iProzd]

Hi @hyeon4977 ,
Both 1 and 2 seem good to me, as long as the dataset itself has correct information and mapping between its own type.raw and type_map.raw files.
For 3, dpdata will rotate the box to a lower triangular in default, see the tag in VASP OUTCAR here: https://github.com/deepmodeling/dpdata/blob/devel/dpdata/plugins/vasp.py#L85 and the function here:
https://github.com/deepmodeling/dpdata/blob/devel/dpdata/system.py#L722 . Maybe you can also do a double-check for some samples to confirm this. :)