Abstract
Predicting full-field mechanical responses accurately and efficiently is of fundamental importance to assess materials failure and has various applications in design optimization, uncertainty quantification, and structural health monitoring. The classical finite element method or global-local scheme can be costly, especially for nonlinear plasticity and damage problems. On the other hand, the homogenization method is efficient for overall mean-field results but fails to capture local full-field responses, which can be critical for materials failure. In this study, we developed deep learning methods to predict full-field plastic reactions, particularly the complicated nonlinear localized plastic shear band patterns in nanocomposites. A Montel Carlo algorithm automatically generates random geometries representing material microstructures of Al/SiC nanocomposites. The models are subjected to pure shear loading of macroscopically uniform boundary conditions admitted by the Hill-Mandel condition in micromechanics. Nonlinear elastoplastic simulations were performed in commercial finite element software ABAQUS to generate inhomogeneous full-field stress/strain responses for data collection and validation. A systematic workflow was created to automate the model generation, finite element simulations, postprocessing, and data curation of response field images for machine learning. After that, a deep learning model of conditional Generative Adversarial Neural Network (cGAN) was developed to predict the full-field plastic response and especially capture the localized plastic shear band patterns. A robust training data set augmented with cases under various rotation transformations is implemented to consolidate unbiased training and ensure the symmetry/objectivity of deep learning models. The proposed image-based deep learning methods can be valuable to researchers deploying data-driven models in many other engineering applications involving large-scale nonlinear full-field predictions.