Development of reduced-order models for predicting the plastic deformation of metals employing material knowledge systems.

Abstract

Metal alloys being explored for structural applications exhibit a complex polycrystalline internal structure that intrinsically spans multiple length-scales. Therefore, rational design efforts for such alloys require a multiscale modeling framework capable of adequately incorporating the appropriate physics that control/drive the plastic deformation at the different length scales when modeling the overall plastic response of the alloy. The establishment of the desired multiscale modeling frameworks requires the development of low-computational cost, non-iterative, frameworks capable of accurately localizing the anisotropic plastic response of polycrystalline microstructures. This dissertation addresses the outlined needs by defining suitable extensions to the scale-bridging, data-driven Material Knowledge System Framework. The extensions detailed in the subsequent chapters enabled the first successful implementation of this framework for predicting the plastic response of polycrystalline microstructures caused by any arbitrary periodic boundary condition imposed at the macroscale. The case studies presented in this work demonstrate that the localization models developed using the MKS framework are of low-computational cost and non-iterative. Nevertheless, their predictions are not as accurate as desired. As a result, leveraging the insights obtained from the implementation of this framework to polycrystalline plasticity, this dissertation provides a robust protocol to incorporate deep learning approaches in order to provide better predictions of the local plastic response in polycrystalline RVEs. The final case study performed in this dissertation establishes that the most robust approaches to develop accurate localization reduced-order models capable of accurately predicting the local anisotropic plastic response of polycrystalline microstructures are deep learning approaches such as Convolutional Neural Networks.