Machine learning streamlines electronic structure calculations for molecules

A machine learning approach could allow computers to determine the electronic structure of molecules without having to use the most resource-intensive equations of density functional theory, new research suggests (Nat. Commun. 2017, DOI: 10.1038/s41467-017-00839-3).

Spam filtering, economic forecasting, and other activities, such as predicting material properties, are powered by algorithms that allow computers to learn from and make predictions on the basis of collections of data. In the case of material properties, computers currently make predictions after being trained with a database of electronic structure information for many types of substances.

A team led by Kieron Burke of the University of California, Irvine; Klaus-Robert Müller of Technical University of Berlin; and Mark E. Tuckerman of New York University now proposes training computers to connect molecules’ structure and properties by having the machines learn from maps of molecular electron density determined from molecules’ potential energy.

After training on existing maps, computers could predict a new molecule’s ground-state electron density. The molecule’s ground-state properties can then be extrapolated from its electron density, as theorized by Pierre Hohenberg and Walter Kohn in the 1960s. In the new work, the team was able to use its “machine learning Hohenberg-Kohn” approach to simulate an intramolecular proton transfer in the enol form of malonaldehyde [CH₂(CHO)₂].