A chemical bond-based machine learning model for dipole moment: Application to dielectric properties of liquid methanol and ethanol

TL;DR

This paper presents a machine learning approach that predicts molecular liquid dipole moments by modeling Wannier centers, enabling accurate dielectric property calculations for liquids like methanol and ethanol.

Contribution

A novel neural network model predicting Wannier centers for chemical bonds, improving dipole moment and dielectric property predictions in molecular liquids.

Findings

Accurately predicts dipole moments of liquids in close agreement with DFT.

Successfully reproduces dielectric spectra over THz to infrared regions.

Identifies physical origins of THz absorption spectra in methanol.

Abstract

We introduce a versatile machine-learning scheme for predicting dipole moments of molecular liquids to study dielectric properties. We attribute the center of mass of Wannier functions, called Wannier centers, to each chemical bond and create neural network models that predict the Wannier centers for each chemical bond. Application to liquid methanol and ethanol shows that our neural network models successfully predict the dipole moment of various liquid configurations in close agreement with DFT calculations. We show that the dipole moment and dielectric constant in the liquids are greatly enhanced by the polarization of Wannier centers due to local intermolecular interactions. The calculated dielectric spectra agree well with experiments quantitatively over terahertz (THz) to infrared regions. Furthermore, we investigate the physical origin of THz absorption spectra of methanol, confirming the importance of translational and librational motions. Our method is applicable to other molecular liquids and can be widely used to study their dielectric properties.