OC20-Dense LMDB with relaxed positions

1

In connection with my earlier inquiry, is it possible to update the IS2RE LMDB of OC20-Dense by incorporating the relaxed positions similar to the OC20 LMDB? We’re considering using the DFT relaxed positions as well.

2

Hi -

I have gone ahead and updated the download link on the challenge webpage to include relaxed positions as well so you don’t need to parse the ASE trajectories directly.

https://dl.fbaipublicfiles.com/opencatalystproject/data/neurips_2023/oc20dense_is2re_train_v2.tar.gz

Let me know if you have any issues.

3

Thank you very much! Is there any possibility that the oc20-dense can be updated in the near future? :slight_smile:

4

We can look into releasing this. The current LMDBs for OC20-Dense don’t have targets since DFT data may not be available for all initial structures (anomalies, failed DFT calculations, etc.). But we can look into releasing a version of the LMDBs that do have lables. Currently that data exists in the form of ASE trajectories if you haven’t already seen that here.

5

Thanks! Absolutely, I can work with trajectories right now. If Meta decides to release lmdb with relaxed positions, please keep me informed!

6

never mind! I processed traj files on my own!

1 Like