LinguaMate: Language‑Guided Metamaterial Discovery via Symbolic-Driven Latent Optimization

ICLR 2026 Conference Submission21723 Authors

Keywords: Metamaterial Discovery, AI for Science

TL;DR: We present a language-guided framework for metamaterial discovery that combines multi-agent collaboration and symbolic-driven latent optimization, enabling practical metamaterial discovery applications.

Abstract: Metamaterials are microstructured materials whose tailored geometries unlock unusual mechanical responses. Metamaterial discovery aims at identifying novel microstructures towards specific applications, such as transportation, robotics, etc. Traditional knowledge-driven metamaterial discovery methods are computationally expensive. While recent data-driven generative models accelerate design, they demand explicit numerical targets and struggle to understand the language descriptions of a concept or idea that is critical for the early design stage. Conversely, large language models readily understand such language intents but lack geometric awareness and physical constraints. To bridge this gap between language understanding and geometric awareness, we propose L**inguaMate**, an inference-time multi-agent optimization framework that empowers language-guided metamaterial discovery via symbolic-driven latent optimization. By jointly aligning language, geometry, and property spaces, LinguaMate discovers physically valid microstructures that extend beyond the boundaries of existing literature and training data. Extensive experiments demonstrate that LinguaMate (1) improves structural validity by up to 34% in symmetry and nearly 98% in periodicity compared to the strongest generative baselines; (2) achieves about 6–7% higher prompt-guidance scores while maintaining superior diversity relative to advanced reasoning LLMs; (3) qualitative analyses confirm the effectiveness of symbolic logic operators in enabling programmable semantic alignment; and (4) real-world case studies further validate its practical capability in metamaterial discovery. We publish our code in https://anonymous.4open.science/r/LinguaMate-CC6F.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 21723