Probabilistic identification of constituent elastic parameters in SiC_f/SiC_m composites via multimodal data fusion and deep neural surrogates

The multiscale heterogeneous structure of ceramic matrix composites poses significant challenges for accurate performance parameter characterization using traditional measurement techniques. This difficulty in quantifying uncertainty limits in-depth analysis of performance variability. To address this challenge, this work presents a novel Bayesian framework for constituent parameter inference. Built on Bayesian theory, a joint likelihood function rigorously accounting for all uncertainty sources is derived. Leveraging multimodal data from strain field measurements and stress–strain curves, the framework resolves the ill-posedness of inverse inference, while Student’s t-distribution is adopted to model measurement noise to yield noise-robust inference. A multi-task ResUNet network is constructed for fast prediction of multimodal outputs, with Monte Carlo dropout integrated to quantify predictive uncertainty. This surrogate achieves a 9.58 × 10⁴-fold acceleration for individual Bayesian inference runs, rendering the computational cost of the framework tractable. Finally, Markov chain Monte Carlo sampling obtains posterior distributions of anisotropic elastic parameters for the fiber bundle and matrix constituents of SiCf/SiCm composites, validated via posterior predictive sampling. This method provides a robust approach for the precise identification and quantification of multiscale performance parameters of SiCf/SiCm composites, laying a rigorous basis for subsequent mechanical performance evaluation and variability analysis of CMCs.