A deep learning approach for inverse design of gradient mechanical metamaterials

Abstract

Mechanical metamaterials with unique micro-architectures possess excellent physical properties in terms of stiffness, toughness, vibration isolation, and thermal expansion. Meanwhile, meta-structures in organisms or geography operate efficiently under complex service conditions thanks to their heterogeneous and gradient distribution of naturally evolved micro-architectures that are difficult to obtain by forward design. In this paper a multi-network deep learning system that satisfies the different design property requirements of microstructures is proposed, and the network predicts the configuration with 99.09% accuracy. The analogy between color space and mechanical parameter space is used to transform parametric design into pixel matching. The microstructures are prepared by AM (additive manufacturing) and their properties are verified by DIC (Digital Image Correlation) experiments (the property error of the structures was less than 2%). Multiscale inverse design of multifunctional and gradient mechanical metamaterials is realized, with special attention payed to the automatic customization of biomimetic structures. The design flow takes only 2 s and the geometric connectivity between microstructure units is considered to ensure compatibility between adjacent microstructures for AM. The proposed design strategy accelerates the emergence of high-performance structures, and provides a reference for topology optimization design of mechanical metamaterials.