MetaGen: A DSL, Database, and Benchmark for VLM-Assisted Metamaterial Generation

Liane Makatura, Benjamin Tod Jones, Siyuan Bian, Md Tamzeed Islam, KARTHICK PRASAD GUNASEKARAN, Shiv Vitaladevuni, Zheng Du, Wojciech Matusik

Keywords: metamaterials, architected materials, microstructures, domain specific language, benchmark, design assistant, coding assistant

Abstract:

Metamaterials are micro-architected structures whose geometry imparts highly tunable—often counter-intuitive—bulk properties. Due to their geometric complexity and a non-trivial mapping from architecture to behaviour, metamaterial design is difficult and inaccessible for non-experts. Vision-language models (VLMs) offer a promising platform for general, democratized metamaterial design, but no existing metamaterial representations or datasets facilitate the creation or evaluation of such an Assistant.
We address these challenges with three complementary contributions. (i) MetaDSL: a compact, semantically rich domain-specific language (DSL) that captures diverse metamaterial designs in a form that is tailored for both humans and VLMs. (ii) MetaDB: a curated repository of more than 150,000 parameterized MetaDSL programs together with their derivatives—three-dimensional geometry, multi-view renderings, and simulated elastic properties. (iii) MetaBench: benchmark suites that probe three core capabilities of VLM design assistants: structure reconstruction, property-driven inverse design, and performance prediction.
We use MetaBench to evaluate the performance of three state-of-the-art VLMs and two fine-tuned variants.
When paired with case studies, our evaluations show that this ecosystem provides a strong foundation for VLM-assisted metamaterial design and learned structure–representation–property relationships.

Primary Area: datasets and benchmarks

Submission Number: 21714