A model and data-driven transverse failure criterion for UD-CFRP composites by considering microstructure distribution and material variation

Developing reliable failure criteria for unidirectional carbon fiber-reinforced polymer (UD-CFRP) composites is essential for advancing structural design and safety evaluation. Traditional criteria are often the average failure envelope of CFRP, lacking sensitivity to the microstructure of fiber distribution. The data-driven approach can predict the strength of CFRP directly from microstructure image, but large amounts of data are required. In this work, we propose a model-data-driven method that derives a microstructure-sensitive transverse failure criterion by integrating a physics-based Tsai-Wu predictor and data-driven residual corrector. Using a convolutional neural network, we first predict the uniaxial tensile, compressive, and shear strengths from microstructural images to initialize the Tsai-Wu predictor as the model-driven component. An input convex neural network is then trained to learn the residual between the Tsai-Wu prediction and true failure strength, forming the data-driven correction component. By predicting failure stresses across numerous microstructures, the failure band is constructed, from which the transverse failure envelope is obtained by fitting its boundary. Notably, the transverse failure envelope can be well described by the classical Tsai-Wu formula with a correction factor μ, which enables the physical interpretability and mathematical conciseness of the obtained failure criterion. In addition, by introducing deep transfer learning technology, the method can be easily extended to new materials with only a few additional samples. The proposed method achieves high predictive accuracy and generalizability with reduced training data requirements, providing a practical strategy for establishing data-enhanced predictive models to support composite failure evaluation.