Expanding the time- and length-scale of <i>ab initio</i> molecular dynamics with deep neural network potentials

Deep neural networks (DNNs) have successfully reproduced the ab initio potential energy surface of condensed phase systems at orders of magnitude lower computational cost. The computational efficiency of DNNs allows molecular simulations of large systems for tens of nanoseconds with the accuracy of ab initio electronic structure calculations. In this talk, after a brief discussion of the Deep Potential (DP) method, I will focus on recent applications of DP molecular dynamics (DPMD) to the study of chemical reactions, amorphous materials and vibrational spectroscopies of water. Such studies require long-time sampling and/or large system sizes, both still out-of-reach of ab initio molecular dynamics. I shall also discuss a generalized version of DP that allows us to describe the electric dipole and polarizability of insulating systems and thus simulate the vibrational spectroscopies of large systems fully from first principles. The methods presented here can be readily extended to a variety of condensed-phase systems combining computational efficiency with the accuracy of quantum mechanics. This work was done in collaboration with Linfeng Zhang, Hsin-Yu Ko, Grace Sommers, Roberto Car and Annabella Selloni.