We develop two machine learning approaches, OrbNet and UNiTE, in which solutions from the Schrodinger equation for molecular systems are predicted by deep neural networks using features based on the molecular electronic structure. Symmetries in the molecular electronic structure are integrated through the roto-inversion equivariant formulation of the presented machine learning frameworks. The proposed machine learning approaches are shown to outperform existing methods in terms of learning efficiency and transferability for the prediction of density functional theory (DFT) results while employing low-cost features that are obtained from semiempirical quantum chemistry calculations. For applications to biological and organic molecules, our approaches predict various chemical properties such as energies, forces, geometries, molecular orbital energies, dipole moments, and electron densities within the accuracy of DFT at a computational cost that is thousand-fold or more reduced.

Geometric deep learning with quantum-chemistry-driven representations

ACS Spring 2022

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