Enhancing predictive performance of artificial neural network modelling composite stiffness degradation with virtual samples

This paper proposes a methodology for generating virtual samples of stiffness degradation of composite laminates to facilitate the training of artificial neural network (ANN)-based predictive models. Considering the impact of matrix cracks, delamination, and fiber fracture on the stiffness of laminated panels, a finite element model driven by the physical mechanism of fatigue damage is established for generating virtual samples of fatigue evolution. The stiffness and evolution of virtual samples for [0/90 3 /0/90 3 ] S glass fiber reinforced polymer (GFRP) laminates are obtained. By employing the fatigue evolution prediction method based on β -variational autoencoder ( β -VAE) and neural ordinary differential equation networks (Neural ODE), the proposed approach effectively learns the underlying mechanisms in fatigue evolution data and shows good correlation to experimental observation. The results indicate that physics-driven virtual sample generation significantly enhances the predictive accuracy and robustness of neural network models, broadening their practical applicability.