Deep Learning Approach for Predicting Stress Fields in Composites With Diverse Fiber Volume Fractions

ABSTRACT Predicting the full‐field stress distribution has important significance in analyzing mechanical characteristics of composite materials. Numerical calculations of stress field distribution using finite element analysis (FEA) can call for significant computational effort for microscale geometries. To address this challenge, this paper demonstrates a deep learning (DL) framework for predicting local stress distributions in fiber‐reinforced composites with diverse fiber volume fractions. An adaptive generation algorithm of representative volume element (RVE) microstructures is developed for constructing stochastic RVE with diverse fiber volume fractions, and the corresponding von Mises stress distribution is calculated using the FEA method. The U‐Net framework is developed, whose weights are trained based on the samples with fiber volume fractions of 30%, 40%, and 50%. Then the weights of the DL model are updated based on additional little samples with random fiber volume fractions between 30% and 50%. The structural similarity index (SSIM) and peak signal‐to‐noise ratio (PSNR) are used for qualifying the accuracy of predicted stress field distributions. The predicted results for a series of microscale geometries show that the mean SSIM values for stress field prediction for diverse fiber volume fractions are up to 98.04%, which can illustrate that the proposed DL model can predict the stress field distribution with diverse fiber volume fractions successfully.