On the use of artificial neural networks and micromechanical analysis for prediciting elastic properties of unidirectional composites

Abstract

The composite design industry has a central demand to predict the elastic behavior of composites from their constituent properties and micromechanical information. In this case, the complex architecture of interlaced yarns in woven composites brings about challenges to accurately predict their mechanical behavior. Multiscale computational methods, often based on computational homogenization, have therefore been established to address the complexity in modeling woven composites. But for computational homogenization of woven composites, one needs to consider the microscale mechanical properties at every point inside a mesoscale unit cell. Based on the possible range of microstructural features, a plethora of research exists to generate random distributions of fibers in a microscopic representative volume element (RVE) and predict elastic properties using numerical methods, such as the finite element method [1,2]. But there is still a requirement to observe the whole possible microstructural design space based on any possible loading case and architecture in order to reach a generic model. Recently,