Double generative network (DGNet) pipeline for structure-property relation of digital composites

Abstract

Deep learning's fast and accurate inference between material configurations and properties has been used to design digital composites with superior mechanical properties. However, initial training sets cannot explore a vast design space with astronomical numbers of possible combinations, and most DL methods cannot guarantee predictive power in unobserved domains that may contain optimal configurations. Active learning-based gradual DNN model update schemes were implemented, but this increased computational costs. We propose a single-shot training pipeline to predict stress/strain distribution and stiffness over configuration space far from the initial training set. Predicting a composite's load response requires predicting its stress/strain field. Two autoencoders and a cGAN predict high-resolution stress/strain fields from grid-averaged local fields inferred from binary digital composite configurations. Our pipeline accurately predicts the high-resolution stress/strain field distri-bution in the unseen volume fraction (VF) domain or for composites with configurations very different from the initial training set. The framework can predict high-resolution fields in many physical phenomena and design composite materials for engineering applications.