Generative metamaterials based on large language models

Abstract

Mechanical metamaterials utilize intricate architectural designs to achieve advanced properties beyond those of their bulk counterparts. Existing metamaterial designs often rely on design inspirations and extensive experimental and numerical studies operated by design professionals, which can be time- and resource-consuming and limited in exploring the vast design space. Here, we transform metamaterial design by developing ChatMetamaterials based on large language models, a prompt-based generative metamaterial design engine capable of inventing architecture codes, and conducting reasoning-based diagnostics and evolution for complex metamaterial systems based on simple text prompts or hand-drawn sketches. This approach changes the way metamaterials are designed, and provides new opportunities for high-throughput metamaterial discovery.