Multilayer Deep Neural Network Modeling of Fatigue Crack Growth in Proton Exchange Membrane

This work proposes a novel hybrid multilayer deep neural network to capture the cross-scale interactions between macroscopic crack growth and mesoscopic plastic-zone evolution, enabling accurate and interpretable prediction of frequency-dependent fatigue crack growth in proton exchange membrane (PEM). PEM is one of the most crucial materials for electrochemical devices. However, fatigue-induced mechanical degradation significantly compromises its safety and durability. Unfortunately, the intelligent damage assessment methodologies associated with multiscale fatigue crack growth behavior of PEM are not yet fully understood. To address this, based on the in situ DIC fatigue testing with four loading frequencies, a hybrid deep learning framework that integrates physical insights and time-series modeling is proposed to predict the fatigue crack growth. Results show clear time-dependent fatigue crack growth behavior. With increasing frequency, the macroscale crack growth rate decreases, while the mesoscopic cyclic plastic zone size increases. The proposed approach comprises six components: (1) Data collection and preprocessing, (2) hybrid neural network modeling, (3) Prediction performance evaluation, (4) small sample optimization, (5) generalization verification, and (6) Shapley Additive Explanations (SHAP) analysis. Comparative model evaluations confirm the framework's predictive accuracy. SHAP analysis identifies loading frequency and plastic zone size as critical factors influencing crack evolution.