A physics informed machine learning framework for fatigue life prediction of additively manufactured Ni-based superalloy

Defects in additively manufactured (AMed) materials are one of the main sources leading to mechanical fatigue failure. Therefore, accurately predicting the fatigue life of materials has become crucial. In this paper, in situ fatigue crack growth experiments were conducted on K477 alloy fabricated by laser coaxial powder feeding (LCPF), experimental parameters and defect characteristics were extracted as input parameters. A physics motivated variable reduction strategy and three data augmentation methods (LI, LIGMM, and LIKDE) were employed for dataset preprocessing. Based on traditional machine learning models, this study proposes a Physics-Informed Neural Network (PINN) model based on monotonic constraints and physical constraints for fatigue life prediction. The results show that the PINN model achieves superior predictive performance, with the coefficient of determination (R²) close to 0.98 and the mean squared error (MSE) on the order of 10^-3, together with strong generalization capability. This experiment-computation integration framework provides theoretical guidance and technical reference for improving the fatigue resistance and service reliability of additive manufactured K477 components in high temperature aero engine applications.