A novel framework of neural network for notch fatigue life prediction by integrating self-attention mechanism and implicit physical constraints

Abstract

Notch fatigue life prediction is critical aspect of structural integrity analysis. This study proposes a new neural network framework − Attention Physics-Informed Neural Network (AttPINN), integrating self-attention mechanism and implicit physical constraints with neural networks. The self-attention mechanism was integrated to extract hidden pattern in input data by allocating attention to input features automatically. Implicit physical constraints were introduced to capture nonlinear relationships between input data and fatigue life without any prior knowledge. For the training and validation datasets with five types of notches and a life range of 10³-10⁷ cycles, the AttPINN predictions were within ± 2 scatter bands, while results from conventional machine learning model and traditional fatigue life equations were within ± 6 and even ± 15 scatter bands. Moreover, the prediction accuracies satisfy within the ± 3 scatter bands for test dataset. This study highlights that the self-attention mechanism significantly improves the neural network feature extraction and modeling efficiency. Implicit physical constraints enhance the generalization ability of the model, particularly for fatigue data with various notch geometries and materials. Results show that AttPINN significantly improved the prediction accuracy as well as generalization ability, making it a promising tool for notch fatigue evaluation.