A physics-embedded linear time-series framework for fatigue crack growth prediction of CT specimens under varying stress ratios

Fatigue crack initiation and subsequent growth are among the primary causes of catastrophic failure in engineering components, making accurate crack growth prediction essential for structural integrity assessment. This study proposes a crack growth prediction framework that integrates physical constraints into data driven sequential prediction models. Two prediction strategies are developed. In the first strategy, fracture mechanics constraints are introduced through the loss function, while in the second strategy, a crack evolution law derived from fracture mechanics is directly integrated into the trend prediction structure. Comparative results show that both approaches achieve improved prediction accuracy, stability, and computational efficiency when compared with purely data driven methods and traditional physical models. In addition, the influence of model architecture is examined by comparing a single trend structure with models that include a residual component. The results indicate that the inclusion of a residual structure consistently improves prediction performance. Overall, the proposed framework reduces computational overhead, improves robustness across specimens with different stress ratios, and ensures better consistency with fracture mechanics principles, demonstrating its potential for online structural health monitoring and fatigue crack growth assessment of critical components.