Walking on Clouds: An Exploration Algorithm for Molecular Dynamic Simulators

Abstract

The energy landscape of atomistic molecular dynamic (MD) simulators provides useful information on how molecules behave and interact under different conditions. However, the dynamics trap trajectories into low-energy potential wells, hindering sufficient exploration. While methods have been created to help alleviate this time limitation, they require knowledge of good coarse variables. Here we present an algorithm that couples nonlinear manifold learning techniques, such as diffusion-maps and local principal component analysis, to MD simulators in order to efficiently and time-effectively search landscapes without a prior knowledge of these macroscopic variables. By establishing the low dimensionality of an attracting manifold and geometrically extrapolating outwards, the simulator can be biased towards rare events, effectively moving up gradients and out of potential wells. We examine different variations of the algorithm and demonstrate the best version utilizing cosine diffusion maps on an intuitive, yet illustrative example.