Revealing and predicting fatigue crack nucleation in textured zircaloy-4 via full-field experiments and physics-informed machine learning

This study investigates the texture-dependent initiation and early evolution of fatigue damage in Zircaloy-4 by integrating in situ digital image correlation (DIC), crystal plasticity finite element (CPFE) modelling, and a strain-gradient-informed long short-term memory (LSTM) framework. DIC measurements reveal pronounced anisotropic strain localisation during fatigue crack nucleation: specimens with out-of-plane c-axis orientation (Z-type) develop earlier crack nucleation and higher strain accumulation than in-plane textured specimens (Y-type). CPFE simulations reproduce the magnitude and evolution of plastic strain and demonstrate that geometrically necessary dislocation (GND) plays a critical role in regulating strain accumulation through gradient-driven strengthening mechanisms. Building on these insights, a physics-informed LSTM model incorporating strain-gradient-derived constraints is developed to predict spatiotemporal strain evolution in fatigue. The strain-gradient-informed formulation significantly improves prediction accuracy, particularly in regions of strong strain localisation, and enables stable long-horizon forecasting. Importantly, validation using an independent unseen fatigue dataset demonstrates robust generalisation capability and confirms that the framework captures transferable deformation mechanisms governing fatigue crack nucleation. These results establish a physically interpretable predictive framework linking strain localisation, GND hardening, and machine learning forecasting, providing a new pathway for predictive modelling of fatigue damage in structural alloys.