search

CederGroupHub / chgnet

Pretrained universal neural network potential for charge-informed atomistic modeling https://chgnet.lbl.gov
https://doi.org/10.1038/s42256-023-00716-3
Other
237 stars 63 forks source link

PBE+VDW Data training #205

Closed haidi-ustc closed 4 weeks ago

haidi-ustc commented 1 month ago

Hello CHGNet team,

I have a set of PBE+VDW data that I would like to use for new training with CHGNet. I would appreciate it if you could provide a training script tailored for this type of data.

Additionally, I have questions regarding the handling of atom_ref in the training process:

How should AtomRef be managed or adjusted when training on PBE+VDW data? Are there specific considerations or best practices I should follow to ensure that the model correctly learns from the provided dataset? Are there any recommended preprocessing steps or adjustments needed specifically for PBE+VDW data before training? Thank you in advance for your assistance! Your guidance will be invaluable as I proceed with this task.

Best regards

BowenD-UCB commented 4 weeks ago

Hi There! If your dataset size is larger than 1000 structure, then it's OK to follow the standard fine-tuning procedure. Typically VDW correction does not bring a large elemental energy difference to AtomRef, in such case you can skip refitting the AtomRef.