Prediction of damage evolution in CMCs considering the real microstructures through a deep-learning scheme

Abstract

The real microstructures of ceramic matrix composites (CMCs) play a crucial role in determining their damage behavior. However, considering the real microstructure within the high-fidelity numerical simulation usually leads to expensive computational costs. In this study, an end-to-end deep-learning (DL) framework is proposed to predict the evolution of damage fields for CMCs from their real microstructures, which are characterized through computed tomography (CT). Three sub-networks, including the microstructure processing network (MPN), elastic deformation prediction network (EPN), and damage sequence prediction network (DPN), are used to construct a two-stage DL model. In the first stage, the geometrical characteristics of real microstructure are precisely captured by the MPN with over 92 % precision for the yarns and matrix. In the second stage, the elastic deformation predicted by the EPN is taken as the intermediate variable to motivate the damage prediction of DPN with the MPN-predicted microstructure as input. The damage evolution of real microstructure is finally predicted with a mean relative error of 10.8 % for the primary damage variable fields. The high-damage regions in the microstructure can also be accurately captured with a mean precision of 87.9 %. The proposed model is further validated by the in-situ tensile experiment. The micro-cracks are proven to initiate and propagate in the high-damage regions. Compared with the high-fidelity numerical methods, this DL-based method can predict the damage evolution on the fly, avoiding time-consuming computation and poor convergence during the damage analysis.