Integrated data-driven identification of constituent elastic parameters for 2D woven SiC_f/SiC_m using a hybrid neural networks accelerated inversion framework

SiC fiber-reinforced SiC matrix (SiC_f/SiC_m) composites exhibit complex multiscale architectures and processing parameters. Obtaining in-situ mesoscale constituent properties has long been constrained by methodological and cost-related bottlenecks. This study developed an integrated data-driven inversion method combining differential evolution algorithm with UNet convolutional neural network and deep neural network (DNN), enabling precise identification of constituent elastic properties of fiber bundles and matrix from experimental measurements. Integrated with a progressive damage model, the inverted parameters accurately replicated the nonlinear tensile behavior of SiC_f/SiC_m specimen. The results demonstrated substantially improved agreement with experiments, reducing the deviation by up to 93.04 % compared to literature-reported values, and enabled reconstruction of complete stress-strain curves from single-load-point measurements. A multidimensional error function incorporating full-field strain and macro-equivalent modulus was established, effectively mitigating ill-posedness in single-data-based inversion and enhancing robustness against Gaussian noise. The integration of hybrid neural networks accelerated the inversion process by approximately 2515 times. Moreover, a weight matrix derived from μ-CT images was employed to correct surface strain, thereby integrating meso-structural information into the inversion framework and mitigating the influence of complex strain noise. This work overcomes limitations of conventional testing methods for multiscale composites and provides an efficient, low-cost approach toward building process-structure-property databases for ceramic matrix composites like SiC_f/SiC_m.