Machine learning surrogates for molecular dynamics simulations of soft materials

Abstract

Molecular dynamics (MD) simulations accelerated by high-performance computing (HPC) methods are powerful tools to investigate and extract the microscopic mechanisms characterizing the properties of soft materials such as self-assembled nanoparticles, virus capsids, confined electrolytes, and polymeric fluids. In this paper, we extend the idea developed in our earlier work of integrating machine learning (ML) methods with HPC-accelerated MD simulations of soft materials in order to enhance their predictive power and advance their applications for research and educational activities. Parallelized MD simulations of self-assembling ions in nanoconfinement are employed to demonstrate our approach. We find that an artificial neural network-based regression model successfully learns nearly all the interesting features associated with the output ionic density profiles over a broad range of ionic system parameters. The ML model generates predictions that are in excellent agreement with the results from MD simulations. The inference time associated with the ML model is over a factor of 10,000 smaller than the corresponding parallel MD simulation time. Through this demonstration, we introduce a “machine learning surrogate” for MD simulations of soft-matter systems. We develop and deploy a web application on nanoHUB to realize the advantages associated with the ML surrogate. The results demonstrate that the performance of MD simulations can be further enhanced by using ML, enabling rapid and accurate simulation-driven exploration of the soft material design space.