Query on Training DP for Composite Systems in LAMMPS

[Terrla]

Hello everyone,
I'm working on a complex atomic system simulation using DeepMd-kit. My study involves atomic types and compounds designated as 'ABCD', where 'A, B, C, D' represents different atomic types, and 'ABC' and 'ABCD' are compounds.

I have a few scenarios and would appreciate insights on the accuracy and feasibility of the approaches:

1. In a system 'ABCD', we have trained a DP model to compute properties for compound 'ABC'. How accurate and reliable would this be, considering the trained model includes 'ABCD' but focuses on 'ABC'?

2. In another approach, we have systems 'ABC' and 'D' trained with separate DP models. Using LAMMPS, we can compute for the 'ABCD' system by employing a hybrid force field setting. However, since our training did not include 'D-A' pairs specifically, I'm concerned about the potential inaccuracies in the results.

3. Furthermore, if our training dataset already contains the 'ABCD' system, would add smaller subsystems like 'AB', 'AC', 'BD', etc., enhance the accuracy or aid in the LAMMPS calculations for the 'ABCD' system? Are there any potential benefits or drawbacks in incorporating these smaller subsystems into the training dataset?

Could anyone provide insights or suggestions on the effectiveness of these approaches, particularly for the scenarios involving separate DP models, hybrid force fields in LAMMPS, and the inclusion of smaller subsystems in the training dataset?

Thank you for your help!

[njzjz]

In short, neural network potentials are many-body potentials. It doesn't directly train type-type pairs.

[njzjz]

I don't think AB and AC will be helpful for ABCD.

[wanghan-iapcm]

In a test of the DPA-1 model on the Al-Mg-Cu system, the training on unitary and binary systems would help improving the accuracy in the ternary alloys. See table I in https://arxiv.org/pdf/2208.08236.pdf

[Terrla]

@njzjz thank you for you clarification.

[Terrla]

@wanghan-iapcm, thank you for sharing the study. I appreciate the reference to the DPA-1 model's training methodology. I'll read the details in Table 1 and see how it might apply to my work.

[njzjz]

@wanghan-iapcm I have a comment on Table I. It doesn't show the results of "training ternary, validating ternary". I think this is required to prove that unitary and binary can improve if ternary exists...