Molecular dynamics (MD) simulations provide theoretical insight into the microscopic behavior of materials in condensed phase and enable computational design of new compounds. However, because of the large temporal and spatial scales involved in thermodynamic and kinetic phenomena in materials, atomistic simulations are often computationally unfeasible. Coarse-graining methods allow simulating larger systems, by reducing the dimensionality of the simulation, propagating longer timesteps, and averaging out fast motions. Coarse-graining involves two coupled learning problems; defining the mapping from an all-atom to a reduced representation, and the parametrization of a Hamiltonian over coarse-grained coordinates. In this talk we will describe a recently proposed generative modeling framework based on variational auto-encoders, where we unify the tasks of learning discrete coarse-grained variables, decoding back to atomistic details and parametrizing coarse-grained force fields. In addition, we will describe the use of on active learning loop to train neural potentials. Although interpolating functions based on neural networks have been shown to learn ground-truth potential energy surfaces very effectively, gathering a training dataset that represents the correct thermodynamic ensemble is challenging. In the proposed approach, a preliminary potential is improved by alternating fast sampling of the phase space using MD on the learned surface, with obtaining validation data in an embarrassingly parallel way.
Copyright. All Rights Reserved.