Adding charges for training periodic systems using global embedding approach

[kousi12]

Hello,

This is regarding the global embedding approach which enables the inclusion of charge and spin during training. I have seen the example script for training OMOL foundation model https://github.com/ACEsuit/mace-foundations/blob/main/mace_omol/mace-omol.sh. Is it possible to include charges for training periodic systems like slabs? If yes, what should be the input, partial charges on the atoms or the total charge on the system? I want to train similar systems with different charges. Also, could you share any documentation or literature that discusses the implementation of the global embedding approach?

Thanks in advance
Kousi

[ilyes319]

Hey @kousi12, while you can input partial charges per atom also if you compute them in another way, in the OMOL model we use input the total charge and the total spin. If you want to do that, just use the same input script as the omol for the embedding part. Yes it is possible to include total charges for slabs too.

For the documentation, best is to look at the release page here: https://github.com/ACEsuit/mace/releases/tag/v0.3.14, and have a look at the code in the embedding: https://github.com/ACEsuit/mace/blob/main/mace/modules/embeddings.py

[kousi12]

Thank you, @ilyes319. One more question. Can I use LES module for fine-tuning (both naive and multihead) the foundation models in MACE?(https://github.com/ChengUCB/les_fit/tree/main) Or can it be used for from-scratch training alone?