High-cycle and very-high-cycle fatigue life prediction in additive manufacturing using hybrid physics-informed neural networks

Fatigue failure remains a critical concern in additive manufacturing (AM) due to inherent defects that degrade mechanical properties and significantly reduce fatigue life. Machine Learning (ML) approaches particularly the Physics-Informed Neural Networks (PINNs), have been employed to predict fatigue life involving small datasets. However, these methods often encounter challenges in managing complex loss functions and reconciling data-driven patterns with established physical laws resulting into rigid training. To address these limitations, a Hybrid Physics-Informed Neural Network (HPINN) that combines the pattern recognition capabilities of Artificial Neural Networks (ANNs) with physical constraints derived from Basquin's law, a modified Paris law, and a non-negativity condition all applied as activation functions is developed in this work. The HPINN model integrates partially trained ANN outputs into parallel physics-based layers optimized using Adam with the Mean Squared Error (MSE) criterion as the overall loss function. The model was validated using fatigue datasets from additively manufactured Al-Mg4.5Mn and Ti-6Al-4 V alloys. Comparative analyses with existing PINN and ANN models show that HPINN consistently outperforms in predictive accuracy, with predictions falling within a 2-factor scatter band. It can also be seen from the three indicators of life prediction that the HPINN model has better performance. HPINN demonstrates the highest R² of 0.99, the lower Symmetric Mean Absolute Percentage Error (SMAPE) of 23 % and Transformed Root Mean Squared Error (T_RMSE) of 1.36 compared with other models. These data indicate the effectiveness of HPINN model in handling complex prediction scenarios and explaining experimental variability.