Multipole-enhanced machine-learning dipole moment predictions in non-equilibrium polycyclic aromatic hydrocarbons

Abstract The accurate prediction of dipole moments for non-equilibrium molecular structures remains a challenge. This study uses a graph neural network to assess the value of incorporating higher-order multipoles, specifically, atomic dipoles and molecular quadrupoles, for predicting the dipole moments of neutral polycyclic aromatic hydrocarbons. Results show that these higher-order terms provide minor improvements over a model using only atomic charges. Combining both multipoles degrades performance, indicating information redundancy. We conclude that for neutral, closed-shell covalent systems, even in distorted configurations, multipole-enhanced models offer a slight added value for molecular dipole moment prediction.