Machine Learning for Vibrational Spectroscopic Maps

Maps that relate spectroscopic properties of a vibrational mode and collective solvent coordinates have proven useful in theoretical vibrational spectroscopy of condensed-phase systems. It has been realized that the predictive power of such an approach is limited and there is no clear systematic way to improve its accuracy. Here, we propose an adaptation of Delta-machine-learning methodology that goes beyond the spectroscopic maps. The machine-learning part of our approach combines Gaussian process regression used to generate the data set with an artificial neural network used to predict spectroscopic properties of interest. A specific application to the OH-stretch frequencies and transition dipoles of water is presented. Our method approximates these properties about two times more accurately than the spectroscopic-maps-only-based approach. Our results become useful in the study of vibrational spectroscopy of condensed-phase systems.