Deep learning, deconvolutional neural network inverse design of strut-based lattice metamaterials

Abstract

Machine learning techniques have furnished a new paradigm in the modeling and design of advanced materials, both in the forward prediction of their effective performance and in the inverse identification of designs that meet specific response targets. While numerous architected media with a diverse range of effective mechanical properties have been investigated thus far, the inverse design of beam-based metamaterials with non-uniform inner architectures that emerge as a consequence of evolutionary optimization processes remains a significant challenge. This contribution elaborates a deep learning, deconvolutional neural network based (DCNN) framework which, when combined with a comprehensive parameterization of discrete lattice spaces, enables the inverse engineering of stochastic lattice metamaterials that cover wide mechanical performance spaces. Auxetic, shear soft and stiff, nearly isotropic and highly anisotropic beam-based metamaterial designs are inversely identified, upon a direct request of their desired mechanical performance, without the need of a latent, condensed space representation. The DCNN model is capable of robustly generating beam-based lattice designs with target mechanical attributes that extend beyond those employed in the initial training domain.