MicroCreep-Net: A novel deep learning framework for predicting microstructure-sensitive creep dynamic crack growth

Abstract

This study introduces MicroCreep-Net, a novel deep learning framework designed to predict the microstructure-sensitive dynamic evolution of creep crack growth with high fidelity and real-time capability. Addressing the high computational cost and time‑consuming nature of conventional finite element–based creep simulations, the framework integrates static grain-boundary morphology with time-dependent crack and strain-field evolution. A large spatiotemporal dataset is generated through batch numerical simulations by using continuum damage mechanics and Voronoi-based polycrystalline modelling. MicroCreep-Net combines a U-Net encoder–decoder, ConvLSTM temporal modeling, dual attention modules, and feature-fusion layers to capture the coupled influence of microstructure and creep damage progression. An incremental prediction strategy and autoregressive iteration enable stable long-term forecasting while suppressing error accumulation. Through comprehensive comparative validation of MicroCreep-Net predictions against high-fidelity numerical simulation experiments, the model achieves a mean structural similarity index measure (SSIM) of 0.9982 with the numerical simulation experiments in short-term predictions, maintains an SSIM above 0.94 in long-term predictions, and retains high accuracy at arbitrary damage stages and across different grain sizes, demonstrating strong generalization capability. This work provides an efficient and accurate approach for microstructure-aware creep damage assessment.