Deep Neural Network Simulations of Bulk and Interfacial Systems

Abstract

Computer simulations of disordered bulk and iterfacial systems were performed with first principles-based methods. Comparison between theory and experiment was achieved through calculation of vibrational spectroscopies, free energies, structure factors and chemical equilibrium constants. Focus was given to three systems of wide interest to chemical sciences: water, TiO2 and the TiO2-water interface. For each of these systems we constructed deep neural network interatomic potentials following the deep potential molecular dynamics (DPMD) methodology. The DPMD models reproduce the potential energy surfaces derived from first-principles electronic structure methods at orders of magnitude lower computational cost. With the computational efficiency of DPMD, ns-long large-scale simulations of disordered systems provided the structure of amorphous and molten TiO2 and the extent of water self-ionization reaction in good agreement with experiment. In addition, a method to construct DNN models of polarizability is described and applied to compute the temperature dependent Raman spectra of light and heavy water. DPMD simulations of the TiO2-water interface answered a longstanding question regarding the nature of water adsorption on the TiO2 anatase (101) surface in contact with liquid water. The methodologies described here can be readily applied to simulate other condensed phase systems, and more importantly, to properly connect the atomistic structure and dynamics of molecular simulations with experimental observables.