Neural network-based material modeling

A neural network-based material modeling methodology for engineering materials is developed in this study. With this material modeling methodology, the stress-strain behavior of a material is captured within the distributed weight structure of a multilayer feedforward neural network trained directly on the stress-strain data obtained from experiments. The feasibility of this approach is verified through constructing neural network-based constitutive models of plain concrete in biaxial stress states and in uniaxial cyclic compression. A composite material model simulating the stress-strain behavior of reinforced concrete as a generic composite material in a biaxial stress state is built with experimental data from Vecchio and Collins' tests on reinforced concrete panels in both pure shear and combined shear with normal stresses. An adaptive neural network simulator is developed by implementing a dynamic node creation scheme and a higher order learning algorithm. Representation schemes, network architectures, training and testing methods, stress- and strain-based approaches for material modeling are investigated. An elastic unloading mechanism is studied with a concrete material model in biaxial compression. Main issues concerning the implementation of neural network material models in finite element solution procedures are briefly discussed. The results on the stress-strain relations of a material predicted by a neural network-based model are compared with experimental data. The developed approach shows promise in the constitutive modeling of composite materials.