Physics Informed Modelling of Fatigue Safety Factors

In this work, AI algorithms are utilized to predict fatigue life based on known stress and strain states, representative of a small-data paradigm with limited experimental data. The work also explores variable safety factors of fatigue models, derived from the model uncertainty induced by the available training data. The approach enables adjustment of fatigue safety margins based on model confidence, reducing unnecessary conservatism against failure while maintaining structural reliability. A physics-informed neural network (PINN) model was trained on experimental data with physical principles modelled mathematically, such that physically feasible predictions are made possible outside of the training data. By utilizing these functions, the PINN model mitigates the influence of limiting experimental data. Predictions of the elastic and plastic deformation from the material testing gives an underlying understanding of the material behaviour for fatigue evaluation. Based on the probability intervals from the fatigue model, a variable safety factor can be estimated, which allowed for a smaller safety factor compared to conservative conventional methods.