Genetic algorithm-enhanced hybrid physics-informed neural networks for very high cycle fatigue life prediction

Abstract

This study proposes a genetic algorithm-enhanced hybrid physics-informed neural network (GA-HPINN) model for very high fatigue life (VHCF) prediction. First, the collected 37 experimental data points were expanded to 100 data points through data augmentation technique,and the augmented dataset was used as the training set. Subsequently, the HPINN model was constructed by incorporating the S-N curve physical model, the Mayer model, and the Z-parameter model as soft constraints into the loss function. Furthermore, the GA-HPINN model was developed by employing a genetic algorithm to optimize the weights assigned to the incorporated physical models. The prediction results demonstrate that the GA-HPINN model achieves an RMSE of 0.2553, exhibiting improved physical consistency and prediction accuracy in fatigue life estimation. It outperforms traditional machine learning methods (Random Forest and XGBoost), PINN models incorporating a single physical constraint, and the HPINN model without weight optimization. The proposed GA-HPINN model addresses the shortcomings of existing dependency methods and provides a more stable and effective solution for VHCF life prediction.