A deep learning-based composite design strategy for efficient selection of material and layup sequences from a given database

A machine learning-assisted composite design framework is established in this paper as an effective and efficient way to find feasible or optimal selections of fiber materials and layup stacking orientations to meet the mechanical and non-mechanical requirements. With a given material database and conventional guidelines for layup design, a generative machine learning model using the physical-informed Conditional Generative Adversarial Networks (CGAN) is developed to provide solutions suitable for various design boundaries such as target strength, maximum deformation, minimum thickness and lowest cost. In order to import the physical constraints in the form of inequalities into the CGAN model, a customized loss along with the traditional loss function of CGAN are applied to the generator for generating outputs that are more favorable to our design requirements as well as conforming to our input data pattern. The networks are trained by using a total of 3000 Finite Element Method (FEM) simulation data from randomly generated composite configurations. Following an evaluation of the training performance of the model, this design strategy is applied to three benchmark problems of a composite tube under different loading conditions and design constraints. The results show that the CGAN model can successfully provide a large number of FRP configurations that fit into the scope of mechanical requirements. In the meanwhile, this integrated CGAN model and FEM system is easily extendable to other composite inverse design scenarios.