Machine learning for accelerating the design process of double-double composite structures

Abstract

Current composite design processes go through expensive numerical simulations that can quantitatively describe the detailed complex stress state embedded in the laminate structure. Nevertheless, these processes usually involve many variables derived from traditional legacy QUAD laminates. Furthermore, numerical simulations, such as finite element method, can include large computational costs that often push cost-benefit compromises. Here we propose to accelerate the simulation process of composite structures by making predictions using a data-driven strategy and adopting a novel family of composite laminates, named double-double, which homogenizes the stacking sequence with fewer plies and only two ply-angles. Multiple machine learning methods are applied to predict the displacement field of a 2D wing-shaped double-double composite model under three loading conditions (tension, shear and bending). The results show that ridge regression can make predictions with the highest accuracy of up to 99% and is faster than simulations by three orders of magnitude, also allowing us to efficiently search for the best double-double angles. This machine learning methodology can be a starting point for more sophisticated simulation models, such as thermo-mechanical loads, complex structures, other composite families, and above all it can simplify the optimization process of composite structures.