Aitomistic Updates: One to Rule Them All——A Universal Interatomic Potential Learning across Quantum Chemical Levels

Have you ever wondered how to train a neural network potential if you have several datasets generated at different levels of theory? Our all-in-one learning strategy elegantly solves this problem by simultaneously training a potential on different datasets. The resulting potentials can make predictions targeting different levels of theory.

This strategy was recently published in JCTC, where we used it to train a universal potential OMNI-P1, which can make predictions at both DFT (wB97X/def2-TZVPP) and coupled cluster (CCSD(T)*/CBS) levels. OMNI-P1 is included in the open-source MLatom.

The all-in-one learning strategy is general and can be applied for training on more levels of theory. In many respects, all-in-one learning is a better alternative to transfer learning, which requires several stages of training and yields two different models.

Due to the uniqueness of the OMNI-P1 model, you can choose the target level of theory to make predictions at. Default is CCSD(T)*/CBS as a better level. After you choose the level, you can use the model for typical simulations, such as geometry optimizations or molecular dynamics.

Below is an example for geometry optimizations in a Python script:

initmol = ml.molecule.from_xyz_file('init.xyz')

omnip1 = ml.methods(method='omnip1', level='cc')

geomopt = ml.optimize_geometry(model=omnip1, initial_molecule=initmol)

final_mol = geomopt.optimized_molecule

print('Optimized coordinates:')

print(final_mol.get_xyz_string())

final_mol.write_file_with_xyz_coordinates(filename='final.xyz')

The simulations with OMNI-P1 can be run online on the Aitomistic Hub, as usual. This version only works for the CHNO elements.

Please also let us know what you would like us to improve to make your experiences with MLatom and Aitomistic better. Simply reply to this email or join us on Slack to directly chat with us.