A novel framework to predict transversal and shear parameters of unidirectional composites by combining experimental, numerical and machine learning methods

Abstract This work presents a new method to predict the transversal and shear properties of unidirectional composites (UD) through combining the experimental, numerical and machine learning methods. The experimental results proved the complexity and difficulty of explaining primary factors affecting the mechanical properties of UD. The representative unit cell model was then created to generate 500 virtual samples for machine learning. The results show that the back propagation neural network model (BP) is the most suitable for predicting the mechanical properties of UD, with an accuracy of 98% within a 2% error. The minimum mean square and absolute errors are 1.09E‐3 and 1.15E‐5, respectively. It is proved that the interface has significant influences on all mechanical properties of UD and shear modulus of composite in 12 directions (G c 12 ) of UD is affected by all input parameters through the optimized BP model. Due to the wide coverage of input data, the proposed BP model is universal and can be adopted to predict transversal and shear properties of UD made from different kinds of fibers. Highlights Interface has influences on all parameters of unidirectional composites. Shear properties of unidirectional composites along 12 directions are intricated. Machine learning can predict mechanical properties of unidirectional composites. Specific samples are beneficial to improve the predicted accuracy.