A multi-stage physics-informed neural network for fretting fatigue life prediction of aluminum alloy

As a complex fatigue-wear coupling phenomenon, fretting fatigue is affected by many factors. Proposing an accurate fatigue life prediction model presents a significant challenge. This paper proposes a physics-informed neural network framework (FFLP-PINN) to predict the fretting fatigue life of aluminum alloys. A multi-stage neural network construction strategy was developed to enhance model accuracy and generalization by progressively incorporating factors related to multiaxial fatigue and fretting damage into the training process. To evaluate the performance of FFLP-PINN, several critical plane models and purely data-driven neural networks model were compared. Meanwhile, the matching of different physical models (critical plane models) with the artificial neural network was carefully investigated. The results indicate that FFLP-PINNs outperform both critical plane models and traditional neural networks. The two-stage framework can further improve model performance. However, not all the embeddings of critical plane models provide satisfactory prediction results. The matching of physical models with artificial neural networks also has an important influence on prediction outcomes. This study provides a new methodology for predicting multi-parameter hard-modeling fatigue cases.