PSI - Issue 75

Laurent Gornet et al. / Procedia Structural Integrity 75 (2025) 129–139 Author name / Structural Integrity Procedia (2025)

133

5

Fig. 2. Gornet- fatigue’s model . Fatigue tests on the [+45/-45/90/0] S laminate. The static failure 699 MPa and the fatigue limit 371 MPa are imposed. Parameters c = 8.930e-03, b = 1.681, d = 0.100, S_sat = 699.00 MPa

4.2. Wöhler (S-N) curve: Neural Networks modelling Neural networks offer a flexible alternative by directly learning from experimental data without strict assumptions about the functional form of the S-N relationship. Neural networks can be trained to predict residual stress or material degradation as a function of cycle count, while respecting boundary conditions such as horizontal tangents at fatigue limits or initial maximum stress values. This approach allows a more faithful representation of typical S-N curve features, including fatigue plateaus, slope changes, and asymptotes, while adapting closely to the experimental data of the specific material under study. 5. Application of Neural Networks to Fatigue Modeling of Materials Fatigue in materials, especially composites, is a complex phenomenon arising from progressive degradation under cyclic loading. Characterizing fatigue behavior commonly relies on S-N curves (stress versus number of cycles), which describe the relationship between load amplitude and fatigue life. Traditionally, these curves are modeled using analytical laws such as Basquin, Coffin-Manson, Strohmeyer or the proposed Gornet- fatigue’s equations. However, these classical models often fall short in capturing the diverse responses exhibited by composite materials due to their anisotropic nature, microstructural variability, and complex damage mechanisms. Neural networks offer a flexible alternative by directly learning from experimental data without strict assumptions about the functional form of the S-N relationship. In particular, Physics-Informed Neural Networks (PINNs) incorporate physical constraints and prior knowledge during training, enhancing prediction accuracy and ensuring physical consistency of the results. Within this framework, neural networks can be trained to predict residual stress or material degradation as a function of cycle count, while respecting boundary conditions such as horizontal tangents at fatigue limits or initial maximum stress values. This approach allows a more faithful representation of typical S-N curve features, including fatigue plateaus, slope changes, and asymptotes, while adapting closely to the experimental data of the specific material under study. Therefore, integrating neural networks into fatigue modeling paves the way for more precise predictive tools, which can support the design and durability assessment of composite structures subjected to cyclic loads.