Neural networks-based fatigue life prediction of natural rubber under combined preaging and temperature effects

Abstract

This study investigates the fatigue behavior of carbon black-filled natural rubber under combined effects of thermal preaging and testing temperature. A comprehensive experimental dataset comprising 410 fatigue tests under 36 distinct conditions is analyzed. Fatigue life is predicted as a function of displacement amplitude, preaging temperature and duration, and test temperature. Several modeling approaches are examined, including an analytical semi-empirical model, a conventional artificial neural network (ANN), an Assisted-ANN, and a Physics-Informed Neural Network (PINN). The ANN models are trained using carefully designed training, validation, and testing datasets to ensure objective performance assessment. Fatigue life is modeled in logarithmic space to improve numerical robustness and reduce sensitivity to data scatter. Model performance is evaluated using quantitative metrics such as mean absolute percentage error (MAPE) and mean absolute error (MAE), as well as qualitative assessment of the predicted S–N relationships. Results show that conventional ANNs significantly outperform the analytical model in terms of prediction accuracy but may produce physically inconsistent S–N curves when trained on sparse and highly scattered data. Incorporating physics-based guidance improves robustness. With the availability of sufficient training data, the Assisted-ANN achieves the lowest overall relative prediction errors with improved physical consistency, and minimal implementation and computational effort. The proposed PINN further enforces physical consistency by constraining the local log–log slope of the S–N relationship, rather than relying on explicit fatigue damage laws or baseline model predictions. As a result, the PINN approach provides the most physically consistent predictions and superior robustness under severe data scarcity conditions. Overall, the results demonstrate that hybrid physics-guided ANN approaches offer substantial advantages for fatigue life prediction of natural rubber under complex preaging and temperature effects.