Material model for composites using neural networks

Introduction A materials such as composites are being used in a variety of engineering applications. These composites exhibit complex behaviors such as anisotropy, microcracking, fiber breakage, etc. Constitutive equations are being developed to describe these complex behaviors using some mathematical rules and expressions based on either experimental data or a theory. The constitutive equations describe the relationship between stresses and strains. A new computational paradigm using Artificial Neural Network provides a fundamentally different approach to the derivation and representation of composite material behavior relationships. Neural network (NN) is a paradigm for computation and knowledge representation inspired by the neuronal architecture and operation of the brain.' There have been considerable research efforts in different applications of NN: signal processing, robotics, structural analysis and design, and pattern recognition' to name a few. Other related work in the use of NN for effective modeling of complex, highly nonlinear relationship among data sets can be found in Ref. 7. The resurgence of earlier research in NN has facilitated the development of a totally different approach to the derivation and representation of material behavior. With this new approach, the knowledge of the material's behavior is captured within the connections of a self-organizing NN that has been trained with experimental data. Recently, the stress-strain behavior of concrete material under the plane stress condition was modeled with a back-propagation (BP) neural network. A neural-network-based material model is developed as an alternative to mathematical modeling of composite material behavior. Neural-network-based modeling solutions are better than conventional methods, such as nonlinear regression analysis, etc., for handling unknown data sets, large dimensional data sets, and noisy data. In this Note, the nonlinear stress-strain behavior of (±6) graphite-epoxy laminates under monotonic and cyclic loadings is modeled with a back-propagation neural network. The NN predicted stress-strain behavior is compared to the experimental data for both monotonic and cyclic loadings.