Using Data-Driven Methods to Learn the Properties of Extended Defects from Atomistic Simulations

Deformation and strength of polycrystalline metallic materials are dominated by motion and interaction of extended defects such as dislocations and grain boundaries. Molecular dynamics simulations have been used successfully for many decades to study the properties of these defects and unit mechanisms pertaining to physical processes involving these defects. In this talk, I will discuss the application of machine learning (ML) techniques on molecular dynamics simulation data to learn the properties of dislocations and grain boundaries in metals. I will demonstrate the utility of both supervised and unsupervised ML to extract physical insights into structure, mobility, and interaction between these defects. In some cases, we observe that the ML models are able to generalize well based on learning from a limited data set, provided that they are trained using a physically meaningful set of descriptors. Techniques to upscale the derived physical insights into mesoscale models for dislocation mechanics will also be discussed.