3D damage evolution and microstructural-based machine learning model for stiffness prediction in woven composite under cyclic loads

Abstract

Tension-compression cyclic loading poses significant challenges due to its severe impact on the stiffness degradation of composites, making accurate predictions of microstructural damage essential for structural reliability. In this study, microstructural damage in woven composites is analysed using X-ray microscopy under equal and critical stress ratio conditions of tension–compression cyclic loading. Quantified damage data are used to train machine learning models, including support vector regression (SVR), random forest (RF) and neural network (NN). Experimental results revealed that under both stress ratios, perpendicular cracks initiated first, followed by cracks at the weft/warp interface. The degradation in the stiffness was approximately 22.46 % under the equal stress ratio condition and 17.62 % under critical stress ratio condition after 100,000 cycles. Machine learning models demonstrated robust performance, with SVR (average error rate = 0.15 %) and RF (average error rate = 0.13 %) closely aligning with experimental data when trained and tested on their respective stress ratios. Notably, flipped dataset analysis revealed that RF (average error rate = 1.02 %) and NN (average error rate = 1.05 %) models trained on equal stress ratio data effectively predicted critical stress ratio behaviour, showcasing their adaptability. These findings highlight the potential of machine learning-driven approaches for predictive modelling, enabling more efficient material design and optimization under cyclic loading conditions.