Finetuning on a dataset with different DFT settings.

[hmcghee22]

When benchmarking the uma-s-1p1 model against my DFT dataset (task = OC20), I see a nice linear correlation with an energy offset since I used different DFT parameters to the ones used when constructing the OC20 dataset.

When finetuning, is this something I need to take into account? Is there somewhere I need to provide my own atomistic reference energies? Or is this accounted for somewhere in the codebase?

Many thanks.

[mshuaibii]

Great question. This is actually already handled here, where your training data is used to fit new atomistic references for fine-tuning.

[hmcghee22]

Thank you!