A machine learning model to predict molecular dipole moments

The molecular dipole moment — a measure of how positive and negative charges are arranged within a molecule — governs how a molecule interacts with light and other molecules. This fundamental property is often computed using density functional theory (DFT) or state-of-the-art quantum chemical (QC) methods. However, such calculations are expensive, and the predicted values are dependent on the method used.

Veit et al. developed a machine learning (ML) model that accurately and reliably predicts molecular dipole moments with low computational cost. Dubbed MuML (after the Greek symbol μ for the dipole moment), the model accounts for two distinct physicochemical effects: local atomic polarization and long-range charge movement. When applied to a dataset of large and complex molecules, MuML yields results as accurate as those using state-of-the-art QC methods, but at a fraction of the computational cost.

The researchers combined two ML methods — the SOAP kernel and Gaussian process regression — that have been used in fitting potential energy surfaces of molecules and materials. By accounting for local environmental effects on atomic charges and atomic dipoles, MuML is able to predict molecular dipole moments with the accuracy of high-level QC methods on the QM7b dataset, which contains small organic molecules with up to seven first-row atoms plus hydrogen.

For a showcase dataset containing larger and significantly more complex molecules, MuML generates molecular dipole moments with DFT accuracy and a much lower computational cost.

As a next step, the authors hope to extend MuML to treat larger and more complex biomolecules, such as proteins and DNA, as well as condensed-phase systems, such as water and molecular crystals.

Source: “Predicting molecular dipole moments by combining atomic partial charges and atomic dipoles,” by Max Veit, David M. Wilkins, Yang Yang, Robert A. DiStasio Jr., and Michele Ceriotti, Journal of Chemical Physics (2020). The article can be accessed at http://doi.org/10.1063/5.0009106 .