Data driven prediction of fatigue residual stiffness of braided ceramic matrix composites based on Latent-ODE

Abstract

A Latent-ODE method is proposed for predicting fatigue residual stiffness of braided ceramic matrix composites, combining β-Variational Auto-encoder (β-VAE) and Neural Ordinary Differential Equation (Neural ODEs). β-VAE is used to extract and separate latent features of underlying fatigue behavior and Neural ODEs is used to learn the underlying dynamics behavior corresponding to the fatigue residual stiffness evolution mechanism. Firstly, the residual stiffness of two-dimensional C/SiC ceramic matrix composites under room temperature fatigue is predicted. Based on the potential variable interpolation method, the prediction results are in good agreement with the calculation results of the phenomenological model. The fatigue residual stiffness prediction method based on partial data retraining can obtain higher-precision subsequent residual stiffness prediction results only using the initial 10% fatigue residual stiffness data. In addition, the high temperature fatigue residual stiffness of two-dimensional braided ceramic matrix composites is predicted. The law of stiffness degradation can be learned by the neural network from the high temperature data, and realizing the high precision reconstruction of high temperature fatigue residual stiffness. The method based on partial data retraining only uses the first 30% of the residual stiffness data to predict the residual stiffness.