A Physics-Informed Neural Network model for combined high and low cycle fatigue life prediction

Abstract

Turbine blades in aeroengines operate under combined high and low cycle fatigue (CCF) loads. Such a complex loading type makes it challenging to accurately assess CCF life. In response, we introduce an innovative framework that integrates physical models with neural networks (NN), known as a Physics-Informed Neural Network (PINN), to boost the accuracy of CCF life prediction. Life data from 11 types of alloys and full-scale turbine blades, with a sample size of 99, are utilized to train and validate the PINN-based model. The results show that, compared to conventional physics-informed models, the PINN provides significantly higher accuracy in predicting CCF life. Based on a substantial amount of experimental data, all PINN predictions fall within 3 times the dispersion bands, whereas the predictions from conventional physical models fall out of 10 times the dispersion bands. We also thoroughly investigate the key factors influencing the prediction accuracy of the PINN model, including the types of physical models and the machine learning algorithms employed. The present study provides a precise and efficient tool for predicting the CCF life of turbine blades.