Generative design of graded metamaterial arrays for dynamic response modulation

Abstract

Mechanical metamaterials (MMs) are micro-structured systems that have long attracted research as well as application interests due to their exotic dynamic functionalities and properties not seen in ordinary materials. Graded MM arrays further enrich the scope of achievable dynamic behavior by employing non-periodic media. Designing such systems is particularly challenging due to the unclear design-performance relationship as well as the heavy computational burden. In this work, we aim to control the time domain response patterns of finite MM arrays and introduce a data-driven design approach that addresses the current challenges. A high-fidelity reduced order modeling (ROM) method is incorporated with statistical learning approaches to realize data generation and physics validation with minimal effort. A variational autoencoder (VAE) is trained to learn the design-performance relation and is used to retrieve classes of design configurations associated with a desired performance. An example application for impact mitigation is shown, and the designed MM arrays exhibit superior protection performance. The combined ROM-VAE approach presents a systematic toolset for designing graded MM arrays with modulated responses, capable of a broad spectrum of tasks such as fast prototyping, inverse generation, and design principle identification.