How to calculate the MAE of the submitted files

[lamthuy]

How to calculate the MAE of the submitted files

[CompRhys]

1. In the CHGNet CSV file both numerical columns "chgnet_energy" and "e_form_per_atom_chgnet" come from predictions of the model. The first is the total energy of the structure as predicted by CHGNet and the second is the formation energy per atom that is calculated from that prediction and the MP reference state energies. CHGNet was trained on the corrected energies and so there is no need to apply the post-hoc MP compatibility settings.
2. "e_form_per_atom_mp2020_corrected" is the correct target to compare corrected energy predictions against.

   matbench-discovery/matbench_discovery/enums.py

   Line 51 in 9a4fe69

   class Key(LabelEnum):

   provides a key to various labels used. There's quite a lot of careful logic that goes into making sure that we get the right numbers when people use different corrections etc and so our preference is for people to just submit predictions and allow us to generate the results using the code in this repo but we can perhaps think about how to document it slightly more clearly to enable more extensible usecases.

[lamthuy]

I double checked again it more likely that the ground truth is the e_form_per_atom_mp2020_corrected from the wbm_summary file because when I calculate the MAE between the e_form_per_atom_chgnet and the e_form_per_atom_mp2020_corrected it is approximately 0.06 which is the reported MAE of chgnet.
This makes me more confused because according to https://figshare.com/articles/dataset/Materials_Project_Trjectory_MPtrj_Dataset/23713842 chgnet was trained on :

-'energy_per_atom': [eV/atom] corrected energy per atom, this is the energy label used to train CHGNet

rather than:

-'ef_per_atom': [eV/atom] formation energy per atom ?

[lamthuy]

Thank you for the explanation, regarding " the second is the formation energy per atom that is calculated from that prediction and the MP reference state energies", is there MP reference state energies for the WBM test structures available? If it is not available for the test structures, is it available for a validation structure?

[janosh]

is there MP reference state energies for the WBM test structures available

yes, have a look at the e_form_per_atom_mp2020_corrected column in matbench_discovery.data.df_wbm.

i'll go ahead and close this issue. feel free to use the discussion tab for future questions. the issue tracker is mostly for bugs and features