Nequip for orbnet dataset

[gianscarpe]

Dear authors,
We're planning to train Nequip on Orbnet dataset (https://figshare.com/articles/dataset/OrbNet_Denali_Training_Data/14883867), which is shipped as ASE xyz files. First, I noticed you recompute energy through ASE calculator when dealing with ASE dataset (at

nequip/nequip/data/AtomicData.py

Line 406 in 9882562

if atoms.calc is not None:

). Is this the suggested way to proceed, or can we adopt the GT energy provided by the dataset?
Second, I've a general question about the "per_species_rescale_scales" and "per_species_rescale_shifts" parameters. The default value is dataset_forces_rms but, as orbnet provides only energy labels, we changed to dataset_per_atom_total_energy_std and dataset_per_atom_total_energy_mean. Can you briefly explaining what is the purpose of this parameter?
Thank you very much for your time!

[Linux-cpp-lisp]

Hi @gianscarpe ,

Thanks for your interest in our code!

This is a consequence of how ASE reads in energy and force data. It should work fine with your data as long as it is in the standard energy key. What do you mean by GT energy?

we changed to dataset_per_atom_total_energy_std and dataset_per_atom_total_energy_mean

Probably the easiest thing (especially on the latest develop, which you should make sure to pull and use) is to set the scales to null (as those config files do) and the shifts as you did to dataset_per_atom_total_energy_mean or the true DFT isolated atom energies.