ML Property Prediction with ALIGNN¶

Predict 50+ materials properties in seconds using ALIGNN graph neural networks.

Install¶

pip install alignn jarvis-tools agapi

Predict a Single Property¶

from alignn.pretrained import get_prediction
from jarvis.core.atoms import Atoms
from jarvis.db.figshare import get_jid_data

atoms = Atoms.from_dict(get_jid_data("JVASP-1002", "dft_3d")["atoms"])

# Band gap
gap = get_prediction(atoms=atoms, model_name="jv_optb88vdw_bandgap")
print(f"Predicted band gap: {gap} eV")

# Formation energy
ef = get_prediction(atoms=atoms, model_name="jv_formation_energy_peratom")
print(f"Predicted formation energy: {ef} eV/atom")

# Bulk modulus
kv = get_prediction(atoms=atoms, model_name="jv_bulk_modulus_kv")
print(f"Predicted bulk modulus: {kv} GPa")

Via AGAPI Agent¶

from agapi.agents import AGAPIAgent
import os

agent = AGAPIAgent(api_key=os.environ.get("AGAPI_KEY"))
response = agent.query_sync("Predict properties of JVASP-1002 with ALIGNN")
print(response)

Batch Screening¶

from jarvis.db.figshare import data
from alignn.pretrained import get_prediction
from jarvis.core.atoms import Atoms

dft = data("dft_3d")
for entry in dft[:20]:
    atoms = Atoms.from_dict(entry["atoms"])
    pred = get_prediction(atoms=atoms, model_name="jv_formation_energy_peratom")
    dft_val = entry.get("formation_energy_peratom", "na")
    print(f"{entry['jid']}: {entry['formula']:8s} ALIGNN={pred:.3f} DFT={dft_val}")

Via AGAPI REST API¶

from agapi.agents.functions import alignn_predict

r = alignn_predict(jid="JVASP-1002", api_client=client)
print(r)

Available Models¶

Key pre-trained models (50+ total):

Model	Property	Units
`jv_formation_energy_peratom`	Formation energy	eV/atom
`jv_optb88vdw_bandgap`	Band gap (OptB88vdW)	eV
`jv_mbj_bandgap`	Band gap (mBJ)	eV
`jv_bulk_modulus_kv`	Bulk modulus (Voigt)	GPa
`jv_shear_modulus_gv`	Shear modulus (Voigt)	GPa
`jv_ehull`	Energy above hull	eV/atom
`jv_magmom_oszicar`	Magnetic moment	μB
`jv_spillage`	Topological spillage	—
`jv_slme`	SLME efficiency	%
`jv_n_Seebeck`	n-type Seebeck	μV/K

Reference¶

K. Choudhary et al., npj Comp. Mat. 7, 1 (2021) — DOI