The history of computational chemistry is the history of finding cost efficient approximations for many-body quantum problem. Since its genesis, computational chemistry used empirical models, parameterizing either parts of the equations needed for complete solution (PPP, AM1, DFTB, etc), an unknown part of the equations (most of DFT functionals), correction to deficient solution (e.g. dispersion correction), or even introducing a model of molecular or atomic system with no reference to electronic structure (molecular mechanics). Recent years, the focus in developments of parametric approximations shifts towards machine learning methods, such as neural networks (NN). The paradigm of NN potentials might be seem close to molecular mechanics methods: it requires ab initio data for parametrization and the result is applicable to same (or very close) domain it was trained on. In contrast to mechanical models, NN potentials might have much broader application domain. It was demonstrated with ANI-1 potential, which is capable to predict chemically accurate DFT energies and geometries of of organic molecules, including reaction barriers and transition states. We present NN potential, which is based upon ANI network. The main aim for developing new version was to make the NN architecture flexible enough to be easily learned (parametrized) towards new molecular properties, new chemical elements, or towards higher quantum chemical levels of theory. Within the network, we have learned representation of atomic properties, atom pair interaction profiles, and atomic representation similar to an “Atom In Molecule” approach. This modular approach enables easy learning of target atomic property, such as energy, charge or polarizability. Along with design of the AIMNet potential, we present case studies of the network performance, generalization and learning new data.

AIMNet: Accurate universal neural network potential for atomic energies, charges and polarizabilities

255th National Meeting (2018)

The official archive and repository for all ACS Meetings & Expositions abstracts and select posters or presentations.