Organizing Data for Training from NEB Results #111
Unanswered
mstapelberg
asked this question in
Q&A
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
-
Hello,
I am trying to train a CHGNet model from NEB results (Energy, Force, Stress). I find that my energy_per_atom MAE is around 2 meV / atom when I assign an mp_id for each ionic step (around 25000 mp_ids). We can call this the Uncorrelated Training.
However when I give 1 mp_id for each image in my NEB calculations, the energy_per_atom MAE goes up to around 80 meV / atom. We can call this the Correlated Training set.
The forces for Uncorrelated are around 0.2-0.3 eV/angstrom and the forces for the Correlated Training set are about 0.3-0.4 eV/angstrom, but more importantly the R^2 values are (E = -0.015, F = -0.007, S = 0.956) for Correlated and (E = 0.999 , F = 0.551, S = 0.998) for the Uncorrelated data
Intuitively, It seems more correct to assign 1 mp_id per image, as the ionic steps should be correlated with one another for each image, but it seems that I get an order of magnitude improvement in accuracy assigning 1 mp_id to each ionic step.
Would appreciate any feedback/suggestions on if I should try and generate more data for the Correlated Training Set, or just use the Uncorrelated Training Set. Happy to provide more information as needed.
Thanks,
Myles
Beta Was this translation helpful? Give feedback.
All reactions