A self-heating-based physics-guided neural network for fatigue life prediction of polymer matrix composites

Great efforts have been made to obtain accurate and reliable fatigue life models for polymer matrix composites (PMCs) that encompass their complexity in a short time frame. This investigation employs a physics-guided neural network (PGNN) with self-heating behavior embedded as soft physics constraints to predict the fatigue life of different PMCs reported in the literature. Due to the limited data size, which ranges from 12 to 21 samples, Monte-Carlo simulations were used to generate synthetic data (up to 1000 samples) that adhere to the distribution of the experimental data. Comprehensive k-fold cross-validation is performed by varying k from 2 up to leave-one-out (LOOCV) on the proposed PGNN architecture to ensure robust prediction and generalization. This model was applied to four distinct PMCs subjected to cyclic loading under stress ratios (R = -1 and R = 0.1) and cyclic frequencies (5, 10 and 50 Hz). The results demonstrate that stabilized surface temperature can predict fatigue life within a factor of two, maintaining reasonable accuracy even for materials that exhibit scatter up to a factor of four. Ultimately, this framework can successfully guide early-stage material screening phase for structural design of PMCs, even when the available experimental data is scarce.