METASYMBO: Multi-Agent Language-Guided Metamaterial Discovery via Symbolic Latent Evolution

Abstract: Metamaterial discovery seeks microstructured materials whose geometry induces targeted mechanical behavior. Existing inverse-design methods can efficiently generate candidates, but they typically require explicit numerical property targets and are less suitable for early-stage exploration, where researchers often begin with incomplete constraints and qualitative intents expressed in natural language. LLMs can interpret such intents, but they lack geometric awareness and physical property validity. To address this gap, we propose MetaSymbO, a multi-agent framework for language-guided Metamaterial discovery via Symbolic-driven latent evOlution. Specifically, MetaSymbO contains three agents: a Designer that interprets free-form design intents and retrieves a semantically consistent scaffold, a Generator that synthesizes candidate microstructures in a disentangled latent space, and a Supervisor that provides fast property-aware feedback for iterative refinement. To move beyond the limitations of reproducing known samples from literature and training data, we further introduce symbolic-driven latent evolution, which applies programmable operators over disentangled latent factors to compose, modify, and refine structures at inference time. Extensive experiments demonstrate that (i) MetaSymbO improves structural validity by up to 34% in symmetry and nearly 98% in periodicity compared to state-of-the-art baselines; (ii) MetaSymbO achieves about 6-7% higher language-guidance scores while maintaining superior structure novelty compared to advanced reasoning LLMs; (iii) qualitative analyses confirm the effectiveness of symbolic logic operators in enabling programmable semantic alignment; and (iv) realworld case studies on auxetic, high-stiffness metamaterial design further validate its practical capability.

• The paper introduces a multi-agent framework that translates free-form language prompts into physically valid, printable metamaterial designs.
• It employs symbolic-driven latent evolution on disentangled latent spaces to achieve high structural symmetry (91.31%) and periodicity (98.35%) compared to baselines.
• Empirical evaluations demonstrate robust diversity and low repeat ratio, validating the method’s potential for hypothesis-driven design refinement in materials science.

MetaSymbO: Multi-Agent Language-Guided Metamaterial Discovery via Symbolic Latent Evolution

Motivation and Problem Statement

Metamaterial discovery requires the generation of microstructured materials exhibiting specific, often complex mechanical properties. In practical early-stage workflows, researchers frequently operate under vague or qualitative design intents, typically expressed in natural language, rather than well-defined numerical targets. Current data-driven inverse-design methods—including VAEs, DMs, and GANs—enable candidate generation but are intrinsically limited to explicit numerical conditioning and operate predominantly within known design spaces determined by training data or literature repositories. LLMs can parse conceptual prompts, but lack geometric awareness or the capacity to enforce structural validity, often resulting in designs that are physically infeasible.

MetaSymbO addresses two central challenges:

• C1: Bridging the modality gap among language (qualitative intent), geometry (microstructural design), and properties (targeted mechanical responses).
• C2: Enabling systematic exploration beyond the confines of known samples and training data to foster hypothesis expansion.

Framework Overview

MetaSymbO is a multi-agent system that orchestrates domain-specialized agents for iterative, language-guided metamaterial discovery. The framework embodies a closed-loop workflow that translates free-form language prompts into validated, printable structures through disentangled latent evolution and symbolic manipulation.

Figure 1: Overview of MetaSymbO. Agent Designer translates the prompt into a scaffold, Agent Generator refines the design in latent geometric space via symbolic-driven latent evolution, and Agent Supervisor evaluates properties to provide fast iterative feedback.

The three principal agents are:

• Agent Designer (Language Modality): Employs LLMs to interpret free-form design intents and retrieve simple, semantically appropriate structural scaffolds by leveraging embeddings and literature. This disambiguates qualitative objectives (e.g., "high stiffness") and anchors them in geometrically meaningful forms.
• Agent Generator (Geometry Modality): Utilizes a disentangled latent generative model to manipulate, compose, and extrapolate geometry beyond the scaffold and existing dataset. This agent features a programmable latent space with explicit factors for lattice, node positions, edges, and semantic properties, supporting symbolic operations including union, mix, intersection, and negation.
• Agent Supervisor (Property Modality): Integrates a trained property predictor and an LLM-based evaluator to provide rapid, property-aware feedback. This agent approximates mechanical responses and constructs alignment scores to iteratively refine both scaffold and generated structures in light of the design intent.

Symbolic-Driven Latent Evolution

A hallmark of the framework is the formalization and operationalization of symbolic-driven latent evolution within the Generator. The latent variable is decomposed into $\mathbf{z}_l$ (lattice), $\mathbf{z}_p$ (positions), $\mathbf{z}_e$ (edges), and $\mathbf{z}_s$ (semantics/properties), allowing for highly controllable inference. Rather than direct decoding of LLM outputs, evolutionary symbolic operators (union, mix, intersection, negation) are enacted over the disentangled subspaces using gradient-based optimization anchored in Sinkhorn-based soft-matching.

• Union: Soft node-level fusion of scaffold and initialization, supporting the introduction of critical structural motifs.
• Mix: Weighted interpolation in latent space, enabling smooth, controllable semantic interpolation between initialization and scaffold.
• Intersection/Negation: Latent operations to extract common semantics or suppress scaffold-dominated features, respectively.

These operators, executed via closed-form manipulation of Gaussians in latent space, enable flexible extrapolation and programmable semantic alignment, as shown in the qualitative operator analysis.

Figure 2: Qualitative analysis shows the proposed operators and latent evolution methods can successfully program the initialized structure towards the semantic of scaffold.

Collaboration and Optimization Loop

MetaSymbO’s collaborative protocol consists of nested Designer–Supervisor and Generator–Supervisor loops. Starting with a user prompt, Designer retrieves a semantically relevant scaffold, which is assessed by Supervisor considering predicted properties and semantic fidelity. This process iteratively refines the input prompt and scaffold until satisfactory alignment is reached, after which Generator conducts guided synthesis and symbolic evolution. The structure is continually evaluated and refined until reaching property and alignment thresholds.

Figure 3: Case study with MetaSymbO. FE simulation denotes finite-element simulation.

Empirical Results

Quantitative Comparison

MetaSymbO delivers significant gains over both state-of-the-art geometric generative models and LLM baselines:

• Symmetry and Periodicity: MetaSymbO achieves 91.31% symmetry and 98.35% periodicity (GPT-4o-mini, Union operator), versus the best generative model baseline with 57.03% and 0.40%, and the best LLM baseline with 85.5% and 65.3%. These results underscore greatly improved structural validity.
• Diversity: Coverage recall exceeds 98% with a repeat ratio below 11%, indicating robust exploration of the structural space without simply reproducing seen samples. LLMs and generative baselines exhibit either high repetition or low periodicity.
• Prompt Guidance: Language-guidance scores are 0.55 for MetaSymbO (GPT-4o-mini, Union), clearly exceeding all baseline LLMs (max 0.50).

  Figure 4: Prediction results of the proposed disentangled VAE on three properties. All $R^2 > 0.8$ demonstrate strong prediction capability.

Case Studies and Real-World Validation

MetaSymbO is applied to prompts such as high-stiffness auxetics, demonstrating closed-loop, multimodal optimization resulting in printable, property-validated designs. Structures are validated via both simulation and 3D printing, corroborating the practical soundness of synthesized metamaterials.

Figure 5: 3D printing and simulation example.

Implications and Future Directions

MetaSymbO establishes a scalable paradigm for language-guided scientific discovery in domains defined by complex, high-dimensional design spaces and weak initial supervision. By integrating symbolic operations over disentangled latents with multi-agent collaboration, the approach bridges the abstraction gap between language and geometry and enables interpretable, modular, and programmable design refinement. Practically, this facilitates hypothesis-driven exploration in materials science and extensible adaptation to other scientific design tasks (e.g., soft robotics, photonic crystal engineering).

On a theoretical level, the latent disentanglement strategy—combined with symbolic Gaussian arithmetic—offers a blueprint for flexible, semantics-aware generation that remains anchored to physically meaningful manifolds. The success of agentic collaboration further supports the integration of LLMs as domain intermediaries rather than monolithic generators.

Future research trajectories include:

• Generalizing symbolic latent evolution to more complex or hierarchical material systems
• Integration of simulation-in-the-loop or active experimental feedback for truly autonomous discovery
• Extension to other property modalities (thermal, optical)
• Unifying prompt-based, property-based, and example-based queries in a single interface

Conclusion

MetaSymbO operationalizes the vision of a language-driven, multi-agent scientist for metamaterial discovery, marrying interpretability, controllability, and practical utility. The framework’s coordinated reasoning over language, geometric, and property domains—anchored by symbolic-driven latent evolution—delivers strong empirical performance on validity, diversity, and semantic guidance, supporting both practical deployment and theoretical innovation in automated materials design.