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

Liane MakaturaMIT CSAIL, Benjamin JonesMIT CSAILSiyuan BianShanghai Jiao Tong UniversityWojciech MatusikMIT CSAIL · 2025

Read MetaGen: A DSL, Database, and Benchmark for VLM-Assisted Metamaterial Generation by Liane MakaturaMIT CSAIL, Benjamin JonesMIT CSAILSiyuan BianShanghai Jiao Tong UniversityWojciech MatusikMIT CSAIL (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-17568.txt.

Introduction

Liane Makatura and co-authors built MetaGen, a tool set for designing metamaterials, and shared it in 2025. Such materials are tiny built structures whose shape gives them odd, tunable traits. Designing them is hard, since shape maps to behavior in ways that are tough to guess. The team aims to let vision-language models help with the task.

The set has three parts. One is a compact code-like tongue for writing designs in a form both people and machines can read. The next is a store of more than 150,000 designs with their shapes, views, and test results. The last is a bench of tests for rebuilding a design, designing to a target trait, and guessing traits. Trained models set the first scores, and testers can try the tools in a design app.

Catalog record

CREATORLiane MakaturaMIT CSAIL, Benjamin JonesMIT CSAILSiyuan BianShanghai Jiao Tong UniversityWojciech MatusikMIT CSAIL
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE116 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-17568

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

MetaGen: A DSL, Database, and Benchmark for VLM-Assisted Metamaterial Generation Liane MakaturaMIT CSAIL, Benjamin JonesMIT CSAILSiyuan BianShanghai Jiao Tong UniversityWojciech MatusikMIT CSAIL 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 150 000 150\,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 baselines by fine-tuning state-of-the-art vision–language models and deploy an omni-model within an interactive, CAD-like interface. Case studies show that our framework provides a strong first step toward integrated design and understanding of structure–representation–property relationships. 1 Introduction Metamaterials have attracted intense research interest because microscale geometries can endow bulk matter with properties that are unattainable in the parent substance. Careful geometric tuning yields extraordinary behaviours such as programmable deformation ( Jenett et al. 2020 ; Babaee et al. 2013 ) , extreme strength-to-weight ratios ( Qin et al. 2017 ) , and simultaneous stiffness and stretchability ( Surjadi et al. 2025 ) . These features enable applications ranging from thermal management ( Fan et al. 2022 ; Attarzadeh et al. 2022 ) to biomedical implants ( Ataee et al. 2018 ; Ambu and Morabito 2019 ) . Yet the design space is effectively limitless, and its full potential remains unexplored. Metamaterial discovery typically follows two paradigms: forward design , which proposes a structure and then measures its properties, and inverse design , which begins with target properties and then searches for a matching structure. Both workflows demand (i) domain expertise, (ii) a grasp of relevant material metrics, (iii) concise yet expressive geometric representations, and (iv) algorithms that map between structural and functional spaces. Vision–language models (VLMs) are well suited to this challenge, as they excel at the cross-modal reasoning, retrieval, and generation required for effective metamaterial design – spanning text, images, 3-D geometry, and numerical property vectors. The complex, verifiable data given by metamaterial design tasks also offers an ideal sandbox for VLM and AI research targeting real-world applications. Despite this symbiotic potential, data-driven metamaterial design is hindered by several issues. For example, Surjadi and Portela 2025 cite the need for “universal tools capable of parametrizing varied architected material morphologies." There is also an acute need for reusable, reconfigurable, task-agnostic datasets featuring diverse structure architectures ( Lee et al. 2024 ) . To address these gaps and catalyse progress in both communities, we introduce the first foundational VLM ecosystem for metamaterial design, anchored by three components: 1. MetaDSL : a domain-specific language that captures metamaterials in a structured, compact, and expressive form accessible to both humans and large language models. 2. MetaDB : a database of more than 150\,000 metamaterials, each of which pairs a MetaDSL program with the derived 3-D geometry, rendered images, and simulated properties. 3. MetaBench : benchmark suites that probe three fundamental metamaterial design tasks – structure reconstruction, property-driven inverse design, and performance prediction – using data sampled from MetaDB. To complete our vision, we also use MetaBench to train and evaluate MetaAssist , a VLM assistant baseline and interactive CAD environment that facilitates multi-modal design interactions including language, images, geometry, and MetaDSL code. All four components are designed for extensibility and community contribution, such that they can evolve seamlessly alongside the state of the art in materials science and agentic design. Collectively, our ecosystem provides a coherent, extensible knowledge base for metamaterial design, while laying the foundation for intuitive, efficient human–AI collaboration in architected materials. 2 Background Metamaterials Two long-standing hurdles in metamaterial design are (i) navigating the immense geometric diversity of candidate architectures and (ii) modelling the intricate, often non-linear mapping from geometry to effective properties ( Makatura et al. 2023 ; Lee et al. 2024 ; Xue et al. 2025 ; Surjadi and Portela 2025 ) . Many studies rely on trial-and-error forward design, where experts hand-craft parameterized structures for specific targets ( Jenett et al. 2020 ; Muhammad and Lim 2021 ; Frenzel et al. 2017 ; Meier et al. 2025 ) . Now, data-driven pipelines also provide a more scalable, systematic alternative in the service of inverse design: Panetta et al. 2015 analysed 1205 families of cubic truss lattices, while Abu-Mualla and Huang 2024 expanded to 17\,000 truss structures spanning six crystal lattices. Beyond trusses, high-throughput workflows create thousands of thin-shell architectures including plate lattices ( Sun et al. 2023a ) and TPMS-inspired surfaces ( Xu et al. 2023 ; Liu et al. 2022 ; Yang and Buehler 2022 ; Chan et al. 2020 ) . Because many datasets target a single architecture class (e.g. beams or shells) and a narrow performance metric, they restrict the attainable property gamut and thus the capability of downstream models ( Berger et al. 2017 ; Lee et al. 2024 ) . Recent designs also increasingly blend classes in hybrid or hierarchical forms ( Surjadi et al. 2025 ; Chen et al. 2019 ; White et al. 2021 ) , emphasising the need for representations that span such boundaries. The procedural-graph approach of Makatura et al. 2023 captures diverse geometries but is demonstrated primarily for human-in-the-loop workflows. Voxel and hybrid encodings scale to 140 k–180 k diverse structures ( Yang et al. 2024a ; Xue et al. 2025 ) , but they sacrifice semantic clarity and compactness, which complicates human or agent editing. Such tradeoffs – along with inconsistencies in geometry descriptors, vocabularies, and evaluation protocols – continue to impede dataset reuse and extensibility ( Lee et al. 2024 ) . We close these gaps with a universal metamaterial descriptor (MetaDSL) along with a reconfigurable database of 150\,000 metamaterials (MetaDB). Each MetaDB entry couples a succinct, semantically rich program with derived 3-D geometry, renderings, and simulated properties, enabling consistent comparison and seamless expansion. Programmatic templating further enlarges the design space, and community contributions can grow both MetaDB and the accompanying benchmark suite. Vision–Language Models for Design Large language and vision–language models (VLMs) have recently permeated design tasks, including procedural textures ( Li et al. 2025 ) , 3-D scenes ( Yang et al. 2024b ; Kumaran et al. 2023 ) , mesh generation and editing ( Sun et al. 2023b ; Wang et al. 2024 ; Jones et al. 2025 ; Huang et al. 2024 ; Yamada et al. 2024 ) , interior layouts ( Çelen et al. 2024 ) , sewing-pattern synthesis ( Nakayama et al. 2025 ; Bian et al. 2025 ) , and computer-aided engineering and manufacturing ( Makatura et al. 2024a ; Makatura et al. 2024b ; Choi et al. 2025 ; Yuan et al. 2024 ) . In most cases, code serves as the medium: pretrained models follow instructions, reuse standard patterns, and emit domain-specific scripts (e.g. Blender Python). When tasks demand novel grammars or specialist knowledge, fine-tuning further elevates performance ( Zhou et al. 2025 ) . Our work

Read the complete text