Spring-based mechanical metamaterials with deep-learning-accelerated design

Abstract

Mechanical metamaterials exhibit unique properties that depend on their microstructure and surpass those of their constituent materials. Flexible mechanical metamaterials, in particular, hold significant potential for applications requiring substantial deformations, such as soft robotics and energy absorption. In this study, we proposed a collection of flexible mechanical metamaterials discretely assembled using structural spring elements. These spring elements enhance both flexibility and reversibility, allowing the materials to withstand large deformations. The geometric regularity of the metamaterials enables zero-shot learning, allowing deep learning frameworks to address property prediction and inverse design problems beyond the training dataset. Using a property-prediction model, the effective mechanical properties of these metamaterials can be accurately predicted based on specified design parameters. Furthermore, an inverse-design model enables the direct generation of mechanical metamaterials with desired target properties, even outside the training dataspace, in the range of Young's modulus E ∈ (0, 350) kPa and Poisson's ratio ν ∈ (-0.12, 0.12). The properties of these inversely designed metamaterials are analyzed through finite element method simulations and mechanical testing. The deep learning-accelerated design approach not only streamlines the development process but also provides a framework for advancing metamaterial design, encompassing property prediction and inverse design. • Developed novel, flexible, and resilient mechanical metamaterials using spring element assemblies. • Demonstrated rapid, zero-shot property prediction and inverse design of these metamaterials via deep learning. • Validated deep learning predictions through simulations and experiments.