Fine tune esen_30m_oma.pt with CP2K MOF dataset, and related fine tune yaml/py files

[JujuHuang]

Dear developer team,

We are planning to fine tune esen_30m_oma.pt model with CP2K DFT datasets of MOFs. We would like to have some suggestions and help from you.

1. esen_30m_oma.pt was trained based on VASP_PBE54_U. Regarding to the situation that we only have CP2K package. We want to know is it okay with fine tuning esen_30m_oma model with CP2K_2023.1_PBE_DFTD3(BJ)? Is is good with different theory/package? we are concerning about the different pseudopotential and basis sets of these two DFT software. Do you have good suggestion about fine-tuning the esen_30m_oma.pt model with cp2k datasets of energy and forces?

2. Does the code (create_uma_finetune_dataset.py) prepare the funetune dataset work for preparing cp2k datasets?

3. As esen_30_oma.pt was trained from fairchem-core-v1, it seems it does not work with using the fairchem-core v2 fine tune config file: https://github.com/facebookresearch/fairchem/tree/main/configs/uma/finetune. I tried to look at any fine tune files from fairchem-core v1, but did not find a good configs that I can follow. I really appreciate for publishing and building these good models, like esen_30m_oma and the fine tuned ones of ODAC. I am wondering would you mind sharing the fine tune code files of ODAC models (like esen_sm_odac25_filtered.pt)?

Best wishes,
Ju Huang