Cluster-guided cGAN-based prediction of whole fatigue crack propagation process across diverse metallic materials

Conventional fatigue crack propagation models face notable challenges in feature dimensionality coverage, scale adaptability, and data scarcity. To address the challenge that existing methods frequently show limited generalization when applied to diverse metallic materials, this study proposes a novel prediction framework—cGAN data augmentation-based deep neural network (cGAN-DA-Net)—that aims to improve the accuracy and robustness of fatigue crack propagation prediction across diverse metallic materials. K-means clustering is first used to guide the data generation process of the cGAN, and the resulting synthetic data are compared with those generated by a traditional GAN. The results show that the cGAN-generated fatigue crack propagation data exhibit superior agreement with real experimental data. The enhanced data are then used to train and evaluate multiple models, including conventional data-driven algorithms (random forest, support vector regression, and K-nearest neighbors), their data-augmented counterparts, and the proposed cGAN-DA-Net model. The findings reveal that while data augmentation improves the performance of conventional models, substantial prediction errors still remain. In contrast, the cGAN-DA-Net model achieves the highest prediction accuracy across different crack scales, maintaining all predictions within a ± 2.5 × error band. These results demonstrate that the proposed cGAN-DA-Net model provides a more accurate and robust approach for predicting whole fatigue crack propagation process across diverse metallic materials. Finally, SHapley Additive exPlanations (SHAP) analysis is employed to interpret the model and identify dominant features, including cycle number, applied load, testing temperature, tensile strength, and key alloying elements (e.g., Cu, C, and Si), thereby confirming the physical consistency and feature integration capability of the model.