Finetuning OC20 model with raw DFT energies or adsorption energies?

[ShaunHan]

Hi. I know that all OC20 models predicts the adsorption energies of a catalyst. I now have some raw DFT energies and forces, and I want to finetune the EquifomerV2 model. Should I use the raw DFT energies directly for training or should I convert them to adsorption energies first? Thanks in advance.

[zulissimeta]

Hi @ShaunHan - you have a couple choices, both of which I would expect to work ok:

1. You could write a dataset with total (raw) DFT energies, and fine-tune a checkpoint trained on total energies
   • This is the most general and most useful model after, but note that you should not just check the MAE on total energies, but how well your model works for what you actually care about.
2. You could reference your energies as adsorption energies, and fine-tune a checkpoint on adsorption energies
   • Note that if you have small numerical errors that might cancel (e.g. k-point convergence errors), this scheme usually works better as the adsorption energies are more well-defined than raw energies.

[ShaunHan]

Thanks @zulissimeta . My question is then how can I tell the code whether my data are raw energies or adsorption energies? Is there a keyword for this for finetuning OC20 models?

[github-actions[bot]]

This issue has been marked as stale because it has been open for 30 days with no activity.