An adaptive loss-balancing PIML method with meta-learning and feature transfer for predicting fatigue cracking of viscoelastic polymer

Machine learning (ML) prediction of fatigue cracking in viscoelastic polymer, such as proton exchange membrane (PEM), is a crucial approach for assessing the remaining life of electrochemical devices. However, physical consistency and stable generalization are limited by insufficient mining of fatigue behaviors, imbalanced data-physics contributions, and inconsistencies across heterogeneous datasets. Herein, a ML prediction method is proposed, integrating fatigue tests, meta-learning-based feature selection, diffusion model-based data augmentation, and domain-invariant transfer learning. Results indicated that crack growth increased with temperature and stress ratio but decreased with frequency. Consequently, a multi-factor modified fracture mechanics model was proposed, effectively capturing the physical information. Based on this, a prediction architecture utilizing a Bidirectional Long Short-Term Memory network with an adaptive data-physics loss-balancing mechanism was constructed. Ten critical features, explaining over 95% of the variance, were identified through meta-learning, while synthetic data demonstrated high distributional consistency with real data. Validation confirmed that the proposed method achieved superior accuracy compared with baseline models, attaining a data-physics equilibrium within 100 epochs. Furthermore, SHapley Additive exPlanations (SHAP) analysis further confirmed that the stress intensity factor range is the key feature governing crack growth. This work provides a physically consistent, data-efficient ML approach for accurate prediction of viscoelastic polymer fatigue life.