TL;DR
MetaGen introduces a comprehensive framework with a new language, database, and benchmark suite to facilitate the design, analysis, and understanding of complex metamaterials using vision-language models.
Contribution
It presents MetaDSL, MetaDB, and MetaBench, enabling structured design, large-scale data, and evaluation for VLM-assisted metamaterial development.
Findings
MetaDSL effectively captures diverse metamaterial designs.
MetaDB contains over 150,000 parameterized programs and properties.
Baseline models demonstrate promising capabilities in structure and property tasks.
Abstract
Metamaterials are micro-architected structures whose geometry imparts highly tunable-often counter-intuitive-bulk properties. Yet their design is difficult because of geometric complexity and a non-trivial mapping from architecture to behaviour. We address these challenges with three complementary contributions. (i) MetaDSL: a compact, semantically rich domain-specific language that captures diverse metamaterial designs in a form that is both human-readable and machine-parsable. (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 test three core capabilities of vision-language metamaterial assistants-structure reconstruction, property-driven inverse design, and performance prediction. We establish…
Peer Reviews
Decision·Submitted to ICLR 2026
1. This paper propose MetaDSL, a domain specific language for metamaterial. This can be regarded as a novel and concise way to express metamaterial structures. 2. This paper builds a large scale metamaterial dataset which contains 3D structures and mechanical properties. The dataset is benificial to the metamaterial community.
1. The motivation is not clear. The authors have not clearly explain why such a DSL is needed. Considering that they want to leverage the power of LLMs, existing textual description of metamaterial can also serve (e.g., depicting truss-based metamaterial with node positions and edges in text format). As for representation, the geometrical representation can already satisfy the need of metamaterial research, like voxel-representation. 2. As the author mentioned, there is a hurdle like "navigating
This paper addresses an important and timely problem—how to enable AI-assisted metamaterial design through structured, interpretable, and verifiable representations. Building a unified ecosystem that connects language, geometry, and physical properties is both necessary and impactful for advancing the field. The proposed MetaDSL is particularly practical and well-designed, offering a compact, human- and machine-readable language that effectively bridges symbolic and physical reasoning. The paper
1. In Table 1, the fine-tuned MetaAssist models show substantial gains over their base models and the zero-shot OpenAI o3 baseline. However, it remains unclear whether any other models fine-tuned specifically on metamaterial data exist, and how they would perform on the same MetaBench tasks. Including such baselines would help better understand the reported improvements. 2. Although MetaDSL is presented as a general, extensible representation, the appendix states that “all of our structures hav
This represents a significant amount of work, a thorough understanding of metamaterials, built from prior art in the domain. The explanation seems solid, the work extensive, the presentation is fine. The dataset is significantly larger than the past efforts they cite. This is indeed a novel dataset.
The paper is, with high likelihood, a significant advancement to the field of Metamaterial design. However, its relevance to a ML venue is questionable. 1. There is no exploration of how benchmark drives innovation in multimodal ML. 2. ML experiments are secondary to materials science contributions. 3. The ML content primarily relegated to the appendix, not the main paper. 4. There is limited insight into model failure modes or challenges in ML tasks.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
