The shift value in fine tune the pre-trained model

[liuchongjun7-cell]

Thank you very much for providing the excellent model. When fine-tuning, we used the single-atom energy of each element in vacuum as the shift for fine-tuning, but the resulting model performance seems unsatisfactory. We suspect that the issue may lie with the shift values. We would like to ask how we can correctly obtain the shift value for each element.

[kavanase]

Hi @liuchongjun7-cell, thanks for using our pre-trained models.
The energies of the single-atom elements in vacuum are what we use as the shift values in the pre-trained models (from here).

So when fine-tuning the models, we recommend using the same isolated atom energies as the shift values, but computed with the same settings as your fine-tuning training data, and updating the model shifts accordingly; see https://nequip.readthedocs.io/en/latest/guide/training-techniques/fine_tuning.html#modifying-per-type-energy-shifts-and-scales