Data-Driven Time-Parallelization in the AFM Simulation of Proteins

Molecular dynamics is a popular technique to simulate the behavior of physical systems, with resolution at the atomic scale. One of its limitations is that an enormous computational effort is required to simulate to realistic time spans. Conventional parallelization strategies have limited effectiveness in dealing with this difficulty. We recently introduced a more scalable approach to parallelization, where data from prior, related, simulations are used to parallelize a simulation in the time domain. We demonstrated its effectiveness in nano-mechanics simulations. In this paper, we develop our approach so that it can be used in a soft-matter application involving the atomic force microscopy simulation of proteins. We obtain an order of magnitude improvement in performance when we combine time parallelization with conventional parallelization. The significance of this work lies in demonstrating the promise of data-driven time parallelization in soft-matter applications, which are more challenging than the hard-matter applications considered earlier.

1. Introduction

Molecular Dynamics (MD) finds widespread use in atomistic simulations in Chemistry, Materials, and Biology. For example, MD provides an excellent method to identify individual conformational states of proteins, and transitions between different conformational states. In MD, forces on atoms due to interactions with other atoms are computed using certain empirical force fields. Once forces are computed, Newton's laws of motion are used, almost always with an explicit time integration scheme, to determine the trajectory of the system. The objectives of MD simulations are two-fold: (i) to determine a statistically representative set of conformational states, and (ii) to reproduce the dynamical transitions between these states.