• Physics-informed LSTM predicts anisotropic creep from mesoscale DIC data. • Creep anisotropy arises from texture-dependent slip activation in HCP Zircaloy-4. • Local-SRS and strain-gradient features enhance predictive accuracy. This study investigates the mesoscale creep behaviour of textured Zircaloy-4 at room temperature by integrating in situ digital image correlation (DIC), crystal plasticity finite element modelling, and a physics-informed recurrent neural-network framework. Pronounced creep anisotropy is observed, with specimens whose c-axes are aligned out of plane exhibiting larger creep strain accumulation than those oriented in plane. Resolved shear stress analysis reveals that this contrast originates from texture-dependent slip activity, where the Y-textured specimen activate both prismatic-〈a〉 and pyramidal-〈c + a〉 systems, while the Z-textured specimen primarily deforms through prismatic-〈a〉 slip. To capture these mechanisms, a dual-stage long short-term memory network is developed that embeds physics-informed descriptors, including local strain-rate sensitivity and strain-gradient features, into both texture classification and strain-field prediction. The creep strain-field problem is reformulated as a long-horizon forecasting task, in which early-stage DIC sequences are used to predict far-future strain evolution. Strain-gradient descriptors act as experimentally accessible proxies for evolving microstructural constraints, enabling physical insight to be incorporated into the learning process. The framework enhances interpretability and long-horizon predictive stability, enabling accurate reconstruction of spatiotemporal creep-strain evolution. This study demonstrates that coupling micro-mechanical experiments with physics-informed learning provides a transferable strategy for predicting texture-dependent creep deformation in structural alloys.
Wan et al. (Sun,) studied this question.