Questions about "atomic-energy"

[jtyang-chem]

I have several questions about "atomic-energy" in training parameters

• set_davg_zero: "Set the normalization average to zero. ", "normalization average" of what? Why it has a relation with atom_ener?
• atom_ener: Is it fixed in training or will change with training steps when being specified in "input.json"?
• start_pref_ae, "If both start_pref_atom_ener and limit_pref_atom_ener are set to 0, then the atom_ener will be ignored.":
  If I set the atom_ener in "input.json", not in file "atom_ener.npy", at the same time start/limit_atom_ener are zeros, does my setting of atom_ener still count? What is the difference between settings in "input.json" and in "atom_ener.npy" of "atom_ener"?

[njzjz]

If you set atom_ener that means the atomic energy in the vacuum, the atomic energy where atoms are not in the vacuum will be given by

where is the descriptor when there is only one atom. is a fixed value. Here, for convenience and performance, we assumes by setting set_davg_zero to true in the program.

When you provide atom_ener.npy and add the loss function, it is a way to fit , which is different from .

[jtyang-chem]

Now I think I understand the equation. And it's different from the one in that "End to End" paper. Is this published?

[njzjz]

The feature is a part of work from this paper.