A machine learning based investigation of void distribution effects in ductile failure

The overarching goal of this work is to develop a predictive model of fracture resistance in a context of epistemic uncertainty due to analytically intractable mechanisms in metallic materials. The uncertainty associated with the spatial distribution of defects, such as voids, and its conflicting effects on setting ductility limits has hindered robust predictions for decades. An immediate objective is to explore the capabilities of carefully selected machine learning methods in uncovering fundamental evolution laws for the spatial distribution of defects, such as voids, under the finite deformations of ductile metals. The approach begins with a simple, yet rigorous classification scheme of void dispersions into ordered, random or clustered. Mathematical arguments are well established for deriving evolution equations of the spatial distribution of voids for both random and ordered distributions. Random dispersions possess the fundamental property of persistent isotropy. Ordered dispersions are characterized with maximum induced anisotropy. State of the art modeling frameworks assume one or the other (random versus ordered). However, real dispersions are clustered due to manufacturing methods. Knowledge of evolution laws for clustered dispersions is critically lacking and is analytically intractable. To address this challenge, machine learning methods are employed in conjunction with surrogate models to supply the data needed for training and testing of the models.