Reverse reconstruction of geometry modeling and numerical verification of 2.5D woven composites based on deep learning

Abstract

This paper proposes a reverse reconstruction method to generating accurately and high-effectively the meso-scale geometry modeling of 2.5D woven composites based on the deep learning. The method first segments the yarn from the X-ray computed tomography (Micro-CT) images based on the deep convolutional neural network (DCNN). Then, reconstruct the yarn surface model from segmented yarn images by using the marching cube algorithm. Consequently, the reverse model is generated by outlining the yarn surface model. Moreover, the yarn geometric parameters are analyzed to evaluate the geometric accuracy of the reverse model. Simultaneously, for the validation of the reverse model, the parametric model, the ideal model, and the experimental tests are considered. Where the parametric model and ideal model are established based on the geometric parameters of yarns. The results show that the DCNN is capable of accurately segmenting yarns from Micro-CT images with a global accuracy of 95.08%. The stiffness prediction error of the reverse model is only 0.706%, which is less than the error of the parametric model (3.77%) and much less than the ideal model. The reverse reconstruction method improves the efficiency of geomerty modeling by focusing on actural images rather than statistical parameters.