Data-assisted physics-informed neural network for predicting fatigue life under various strain ratios and pre-strain effect

Hydrogen storage cylinders generally experience repetitive pressurization and depressurization cycles, leading to cyclic stresses that can cause progressive fatigue damage. Accurate fatigue life prediction is essential for ensuring the structural integrity and operational safety of these cylinders. Given the challenges posed by limited experimental data and poor accuracy with small datasets, this study introduces a data-assisted physics-informed neural network (DA-PINN) for low cycle fatigue life prediction under various strain ratios and pre-strain effect. Bootstrapping and spline interpolation techniques are used to enrich both the training and testing data set of the neural network, ensuring a more robust learning process. Simultaneously, physical knowledge of strain-controlled fatigue behavior is embedded into the model through constrained loss functions, enabling physics-consistent learning. Results demonstrate that the proposed DA-PINN achieves prediction accuracy predominantly within a ±2.0-fold error band, outperforming conventional empirical models. Moreover, comparative analysis reveals that the bootstrapping augmentation method yields more consistent predictions than spline interpolation. The developed DA-PINN exhibits strong material adaptability, suggesting promising applications for other metallic materials under various strain ratios and pre-strain effect.