Scale-bridging from the Atoms Up; Employing Machine Learning to Improve the Accuracy and Scalability of Molecular Dynamics

Simulation methods such as Molecular Dynamics(MD) are powerful tools for examining behaviors of materials that originate on the atomic scale. On modern supercomputers it is possible to simulate many millions of atoms on time scales exceeding one microsecond. Fitting an interatomic potential(IAP) is a critical multiscaling link between electronic structure codes, such as density functional theory and MD that determines the accuracy of these simulation efforts. Our approach to constructing this multiscaling link employs machine learning using the Spectral Neighborhood Analysis Potential(SNAP) to learn the features of a DFT database. This work is focused on the tungsten-beryllium system to better understand the effects of beryllium transport to the tungsten diverter in ITER-like conditions, where accurate predictions of implantation depth profiles and surface behavior will be important.Experimentally, a beryllium-tungsten intermetallic forms under fusion relevant conditions, molecular dynamics simulations of both energetic beryllium implantation as well as beryllium deposition on a tungsten surface were performed to investigate this surface chemistry using this newly developed SNAP potential.