Fast prediction of viscoelastic behavior of 3D tubular braided composites based on deep learning

Rapidly and accurately calculating the macroscopic viscoelastic properties of three-dimensional (3D) tubular braided composites is of great practical significance for their structural design and optimization. This study proposes a data-driven approach combined with trans-scale modeling to predict the axial compressive viscoelastic properties of 3D tubular braided composites. First, the viscoelastic constitutive relations for the matrix and the yarn are established, and the trans-scale finite element model of the 3D tubular braided composites is constructed based on micro-CT technology to calculate the viscoelastic curves and perform experimental validation. Then, a deep neural network (DNN) model integrated with an automatic hyperparameter optimization algorithm is built to train and test the simulation dataset generated by the finite element model, and finally the axial compression relaxation modulus curves of 3D tubular braided composites with different parameters (braiding angle, fiber eccentricity, inter-yarn porosity, intra-yarn porosity, temperature, total fiber volume fraction and fiber elastic modulus) are predicted. The results show that the developed data-driven model based on finite element and deep learning can quickly and accurately predict the macroscopic viscoelastic properties of 3D tubular braided composites.