Abstract

The dipole moment of a molecule is vital in understanding its interaction with other molecules and with light (and thus its spectroscopic properties). Despite its importance, an accurate quantum-chemical calculation can be computationally expensive, and sensitive to the computational details. I describe a method for learning molecular dipole moments by combining two machine-learning methods that each represent a different type of physical effect: scalar Gaussian Process Regression (GPR), which learns partial atomic charges and thus resembles charge separation, and vector GPR , which learns partial atomic dipoles and resembles atomic polarization. While these two methods have their own advantages and disadvantages, combining the two gives rise to a single method with the best features of both, opening up the way to accurate molecular simulations and calculations of spectra.