Formation energy calculation

[ricardonpa]

Hello, is there a practical way to calculate the formation energy from the total energy predicted by CHGNet? I mean something like get_e_form_per_atom from matbench_discovery, but this utilizes relative energies from DFT rather than CHGNet. Is there anything analogous that is pre-computed for CHGNet, or any recommended workflow for getting self-consistent formation energies on the CHGNet scale?

from pymatgen.core import Structure
from chgnet.model.model import CHGNet
from matbench_discovery.energy import get_e_form_per_atom, mp_elemental_ref_energies

from urllib.request import urlopen

url = "https://raw.githubusercontent.com/CederGroupHub/chgnet/main/examples/mp-18767-LiMnO2.cif"
cif = urlopen(url).read().decode("utf-8")
structure = Structure.from_str(cif, fmt="cif")

chgnet = CHGNet.load()
prediction = chgnet.predict_structure(structure, task='e')

print(f"Predicted energy: {prediction['e']} eV")

entry = {'energy': prediction['e'], 'composition': structure.composition.element_composition}
formation_energy = get_e_form_per_atom(entry, mp_elemental_ref_energies)
print(f"Formation energy per atom: {formation_energy:.4f} eV/atom")

CHGNet v0.3.0 initialized with 412,525 parameters
CHGNet will run on mps
Predicted energy: -7.367691516876221 eV
Formation energy per atom: 4.3201 eV/atom

[bowen-bd]

You can relax the elemental phases or use their energies from the Materials Project