Neural Network-Based Framework for Fatigue Life Prediction of Natural Rubber Under Thermo-Oxidative Preaging Spectra

Abstract

Thermo-oxidative (TO) preaging significantly influences the fatigue behavior of natural rubber (NR). However, existing artificial neural network (ANN) models are generally limited to isothermal aging conditions and cannot directly represent multi-temperature preaging spectra encountered in practical applications. This study develops a framework for extending ANN-based fatigue-life predictors trained exclusively using isothermal preaging data to multi-temperature aging histories. An optimized physics-informed neural network (PINN) is employed as the underlying fatigue-life predictor, incorporating a physics-based constraint to preserve the expected power-law behavior of the generated S-N curves. Two spectral-aging representations are investigated. The first converts the aging spectrum into an equivalent isothermal condition using weighted-mean-temperature or Arrheniusbased equivalent-time formulations. The second introduces an incremental-degradation representation in which the individual aging steps are evaluated sequentially while accounting for the preceding aging history. The optimized PINN achieves mean absolute percentage errors (MAPE) of 16.5% and 14.8% for unseen data within the training conditions and left-out isothermal aging conditions, respectively. For spectral preaging, the Arrhenius-based equivalent-time method substantially outperforms the weighted-mean-temperature approach. The proposed incremental-degradation representation provides the highest accuracy, with a MAPE of 6.13%, while preserving the sequential nature of the aging process and avoiding the need for a global reference temperature. Its predictions are also in close agreement with the experimental S-N curve and compare favorably with the analytical Neuhaus model, without requiring explicit determination of temperature and displacement dependent activation-energy parameters. The proposed methodology therefore provides a practical approach for extending isothermally trained ANN-based fatigue-life models to multi-temperature preaging histories.