Data-Driven Upscaling—Applications to Multi-Scale Modeling of Extended Defects

Nucleation, propagation, and interaction of extended defects, such as dislocations and grain boundaries, play a critical role in controlling the mechanical response of polycrystalline materials. Multi-scale modeling, primarily based on mechanistic and phenomenological models, has played a key role in understanding these phenomena and predicting material response. Advent of exascale computing and advanced machine learning algorithms opens up new venues of <i>data-driven upscaling</i> where physics-informed machine learning models, in conjunction with physics-based descriptors, can effectively bridge between different computational scales with minimal constitutive and/or phenomenological assumptions. In this talk I will discuss data-driven upscaling applied to modeling of extended defects, primarily focusing on atomistic simulations and discrete dislocation dynamics. It is shown that uncertainty-driven active learning can efficiently learn the underlying physics, where the only assumptions are based on well-grounded physical constructs such as mean-free paths, thermal activation, and configurational forces.