Physics-informed machine learning for high-cycle fatigue of AM Ti6Al4V: Life prediction and correlation analysis

The process-structure–property (PSP) relationship in additive manufacturing (AM) has always being a significant topic, directly governing material optimization and reliable performance prediction. To address this challenge, this paper proposes a novel neuro-Basquin PDE constrained network for life prediction and correlation analysis of AM Ti-6Al-4V. The proposed neural network architecture is fundamentally based on the Basquin equation, yet exhibits nonlinear descriptive capabilities that significantly surpass the original equation, while further incorporating partial differential equation-based constraints into the loss function to guide model training. An inverse configuration that takes life as the input and stress as the output is adopted, ensuring good convergence of the model while accurately describing the fatigue limit of the data. To address parameter incompleteness in some datasets, an XGBoost-based imputation strategy was proposed. The results show that the proposed model can predict fatigue strength and fatigue life very well and has excellent generalization performance.