Author Description

Molecular dynamics simulations accelerated by high-performance computing methods are powerful tools for investigating and extracting the microscopic mechanisms characterizing the properties of a wide range of materials, including soft matter. Generally, the runtimes for these simulations range from hours to days to furnish accurate information using optimal parallelization techniques. These simulations produce enormous amounts of data that can guide the design and engineering of materials and devices. However, these large datasets can be very high-dimensional and thus challenging to be analyzed by conventional processing approaches. There is a critical need for new approaches to accelerate simulations, leverage past simulations to generate accurate predictions, and expedite the analysis of simulation data to investigate and understand material properties. This dissertation contributes to addressing this need. We discuss broad and effective deep learning approaches for training machine learning surrogates and demonstrate their success in molecular dynamics simulations of soft-matter systems. We introduce artificial neural networks based timestep auto-tuners for molecular dynamics simulations and develop recurrent neural networks based integrators that accurately solve Newton's equations utilizing sequences of past trajectory data using timesteps that are orders of magnitude larger than the Verlet timestep. We illustrate how machine learning techniques can perform feature investigation of high-dimensional datasets generated in nonequilibrium molecular dynamics simulations of polymeric liquids. Finally, we demonstrate the use of advanced cyberinfrastructure such as transient cloud servers to reduce the computational costs of molecular dynamics simulations of soft matter.