Rapid evaluation and prediction of cure-induced residual stress of composites based on cGAN deep learning model

In this piece of present work, we propose a deep learning model driven by conditional generative adversarial network (cGAN) for rapid evaluation and prediction of cure-induced residual stress (CRS) of composites. The CRS is evaluated by solving for the material behaviors of the cure kinetics, viscoelasticity, thermal expansion, and cure shrinkage under heat transfer condition using finite element (FE) method. The geometry and CRS fields are extracted from numerous simulations and subsequently utilized in the proposed cGAN training process. With this end-to-end unsupervised learning, the cGAN model can predict the CRS with high fidelity based on the fiber random distributed microstructure and capture the variation of fiber volume fraction. In qualitative measures, peak signal to noise ratio (PSNR) and structure similarity index (SSIM) are employed for accuracy verification. Moreover, the cGAN model can also significantly reduce the computational cost and provide some insights for optimizing the manufacturing process of composites.