A novel artificial neural network-based failure criterion for unidirectional composites under multiaxial stress states

Carbon Fibre Reinforced Polymer (CFRP) composites are widely used in many engineering applications due to their excellent design flexibility and high stiffness- and strength-to-weight ratios. However, the lack of comprehensive experimental data for the validation of computational failure models, especially for composite structures subjected to multiaxial loadings, has led to highly conservative designs. Here, high-fidelity finite element-based 3D representative volume element models are developed to analyse the failure mechanisms of IM7/8552 CFRP unidirectional (UD) composites subjected to biaxial loadings via periodic boundary conditions [1]. The Drucker-Prager plastic damage constitutive model and cohesive zone model are utilised to simulate the mechanical response of the matrix and fibre-matrix interface, respectively. Fibres are assumed to be transversely isotropic and brittle material. Fibres failure is predicted using the maximum principal stress criterion. Due to the transverse isotropy of the cross-section of UD lamina, nine out of fifteen loading stress combinations are selected. For the generation of failure points and representative modes, an average of ten load cases are considered for each selected biaxial stress combination as shown in Fig. 1. For a particular biaxial loading condition, there is main failure mode, which can be assigned to the corresponding loading case. The failure surface of the UD composite was then fitted by using the univariate spline function. The data points within this surface are defined as “safe”, and those beyond it are defined as “failure”. A database with half a million samples was used for training, validation and testing of artificial neural network (ANN). Six stress components are selected as inputs and 0/1 is selected as output (“safe/failure”). The accuracy of ANN can reach 98.5% after training.