Voltage-dependent surface IR spectra of water at gold electrodes from machine-learned ab initio dipoles

Surface vibrational spectroscopy is a powerful tool for characterizing electrochemical interfaces. In particular, when supported by atomistic simulations, it can provide critical information on the roles of structure and bias in electrochemical reactions. However, accurate simulation of these spectra is difficult, given that such systems require a dynamical description of the consequences of applied voltage on the electronic structure of the electrolyte phase. Here, we obtain the surface-specific infrared spectra of water at a gold electrode by combining ab initio molecular dynamics trajectories with deep neural networks trained to represent the dipoles of the interfacial water. This approach was carried out for trajectories representative of five different applied voltage biases. The resulting vibrational spectra yield significant agreement with the experimentally observed effects of both positive and negative applied voltages. The computed spectra also unravel the interplay between the orientation, hydrogen bonding, and dipoles of interfacial water and its key vibrational signatures. This protocol and the results shed light on the behavior of water at electrified interfaces and, moreover, demonstrate the utility of machine learning of ab initio data for computational spectroscopy.