Reliability prediction framework for woven CMCs structures based on multi-scale uncertainty and artificial neural networks

This study addressed the uncertainties in mesoscopic geometric and performance parameters of woven ceramic matrix composites induced during the weaving and densification processes. Based on the concept of multi-scale stochastic propagation, a neural network prediction method for structural reliability was proposed. Fiber tests were conducted to obtain the distributions of fiber modulus, strength, and radius, which were used to simulate the mesoscopic performance distribution of yarns. Mesoscopic geometric parameter distributions were derived from XCT, and the harmony search algorithm was employed to optimize parameter combinations. A dataset was established using multi-scale simulation methods. A two-level ANN reliability prediction surrogate model was developed, achieving an average prediction error of 3.81% for the structural failure load, with the predicted failure regions aligning with experimental results. Sensitivity analysis of mesoscopic parameters based on the PAWN and SHAP methods revealed that the minor-axis length has a significant influence on the structural failure load.