Data-driven PSP linkages for atomistic datasets

Abstract

For a variety of materials, atomic-scale modeling techniques are commonly employed as a means of investigating fundamental properties, including both structural and chemical responses. While force-field based calculations are significantly less computationally expensive than their quantum-mechanical counterparts, the datasets often investigated are large in size (10^3 – 10^9 atoms) and high-dimensional, and thus cumbersome for use in multi-scale models. The development of quantitative “process-structure-property” (PSP) linkages for atomistic simulations presents a powerful route to convert atomistic simulation data into actionable knowledge. Here, a framework is presented for quantifying structure from these simulations in full- and reduced-dimensional form, and a series of protocols are developed for establishing regression models for process-structure and structure-property linkages.