A multi-fidelity data-driven framework for predicting mechanical property distributions of composite structures and its validation

The failure mechanisms of carbon-fiber-reinforced composites are complex, and experimental tests are costly. Traditional finite element methods, limited by current theoretical models, struggle to accurately simulate the entire failure process and exhibit significant cumulative errors, complicating precise modeling and uncertainty quantification. Machine learning approaches offer a promising alternative but generally require extensive datasets to achieve satisfactory performance. We present a multi-fidelity data-driven framework that blends a small set of high-fidelity test results with an extensive collection of low-fidelity simulation data to predict the distribution of mechanical properties in composite structures. The framework is validated through tensile-failure experiments on notched laminates. To improve the statistical representativeness of the limited experimental samples, we introduce a Bayesian data-augmentation scheme and derive the theoretical distribution of the inter-group coefficient of variation to confirm its soundness. Cross-validation shows that the proposed method attains a mean absolute error of 4.99% when predicting the 10th percentile of the failure-load distribution. The study mitigates the twin challenges of scarce experimental data and weak coupling between numerical models and physical tests.