| In this study, the viscoplastic behavior of a carbon-fiber/thixotropic-epoxy matrix composite was investigated through two deferent modeling efforts. The first model is phenomenological in nature and it utilizes the tensile and stress relaxation experiments to predict the creep strain. In the second model, the composite viscoplastic behavior is no longer represented by closed-form constitutive laws, but it is captured by a neural network formulation.; The composite was processed and cured using hand lay-up technique together with autoclave curing system. By performing thermomechanical analysis and differential scanning calorimetry, the glass transition temperature of the composite was noticed to degrade. Experiments were performed to examine the tensile, creep, and load relaxation behavior of the composite under different temperatures.; It was found that the composite strength and stiffness decrease significantly at elevated temperatures. A phenomenological model was constructed based on the overstress viscoplastic model. In this model, four material's parameters are extracted from tensile and load relaxation tests. These parameters are used by a system of coupled equations to predict the creep strain.; The results of the phenomenological model were satisfactory for predicting creep at low temperature conditions, but for the high stress-high temperature regimes, the model failed to predict the creep strain accurately.; The neural network model was built directly from the experimental creep tests performed at various stress-temperature conditions. The optimal structure of the neural network was achieved through the universal approximation theory and the dimensionality of the creep problem (stress, temperature, and time). The neural network model was trained to predict the creep strain based on the stress-temperature-time values. The performance of the neural model is captured by the mean squared error between the neural network prediction and the experimental creep strain results. To minimize this error, several optimization techniques were examined. The minimization of the error when carried out by the scaled conjugate gradient outperforms the standard backpropagation in terms of convergence rate and accuracy. Using neural network with scaled conjugate gradient training algorithm, the prediction of the creep strain was very satisfactory compared to Gates phenomenological model. |
