Application of neural network to model stiffness degradation for composite laminates under cyclic loadings

This paper investigates the applicability of modelling stiffness degradation in fiber reinforced polymer (FRP) composites with a state-of-the-art artificial neural network (ANN) architecture. β-Variational autoencoder (β-VAE) is first applied to extract disentangled latent features to represent the underlying driving mechanism. A neural ordinary differential equation (neural ODE) is then adopted to learn the dynamics of the latent features, which enables a continuous prediction of the stiffness over the cycle-domain. The ANN model is trained and validated before compared to both conventional mechanical and phenomenological models, where the ANN-based model shows comparable performance. In addition, a latent S–N curve is proposed based on latent variable analysis, which shows better correlations to the experimental data over the traditional S–N curve. Overall, with recent rapid developments in ANN architectures and algorithms, the ANN model is found to be a very promising tool for solving fatigue-related engineering problems for FRP structures when properly used.