Deep Potential Molecular Dynamics*

The interfaces of liquid water and ice with air abound on Earth and impact non-negligibly atmospheric sciences. Directly probing the structure of interfacial water is challenging, and the surface-specificity of Sum Frequency Generation (SFG) spectroscopy has offered important insights on both structure and dynamics of interfacial water and ice. Ab initio molecular dynamics (AIMD) could, in principle, aid interpret the experimental observations since it accesses both molecular and electronic structure of interfacial water. However, AIMD stumbles on the slow convergence of the SFG spectrum with simulation time. We here overcome this difficulty using Deep Neural Networks (DNN) to represent the ab initio potential energy surface. DNNs were also used to obtain molecular dipole moment and polarizability of water in agreement with first-principles methods, which allows a proper computation of the SFG spectrum of water/air and ice/air interfaces relying only on the dipole approximation. Our method can be easily extended to other interfaces and should enable ab initio level simulations of SFG spectra for a wide variety of systems.