⚡ electrical engineering

M³Sculptor: A Knowledge-Driven Multimodal-Agent Framework for Automated Metamaterial Design

M³Sculptor is a knowledge-driven multi-agent framework that leverages large language and multimodal models to automate the design of mechanical metamaterials by retrieving literature, reconstructing 3D structures, and iteratively refining them through symmetry restoration and computational homogenization to meet specific natural-language requirements.

Original authors: Sheng Sun, Qing shang, Menghuan Wang, Yang Lu, Qining Wang, Yawen Fei, Tong-Yi Zhang

Published 2026-11-05

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Sheng Sun, Qing shang, Menghuan Wang, Yang Lu, Qining Wang, Yawen Fei, Tong-Yi Zhang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

For centuries, engineers have built stronger, lighter, and more resilient materials by changing what things are made of. They mix metals, blend polymers, or refine ceramics to get the right strength or flexibility. But in the last decade, a different idea has taken hold: that the secret to a material's behavior lies not just in its chemical recipe, but in its shape. Imagine a sponge made of the same plastic as a hard brick. The brick is solid and heavy, but the sponge is light and squishy. Now imagine a material where the internal structure is so precisely engineered that it can stretch sideways when you pull it, or become softer the harder you squeeze it. These are mechanical metamaterials. They are not new substances, but carefully arranged architectures that can do things nature never intended. The problem is that designing them is incredibly difficult. The number of possible shapes is so vast that trying them one by one is like searching for a single grain of sand on a beach without a map.

Researchers have turned to computers to help, but even the smartest programs struggle with the sheer volume of information locked away in scientific papers. For years, scientists have published thousands of studies describing these strange structures, often showing them as two-dimensional drawings. To use this knowledge, a researcher usually has to read the paper, understand the drawing, manually build a three-dimensional model, and then run complex simulations to see if it works. This process is slow, prone to human error, and often stops before the best design is found. A team of scientists has now created a system called M3Sculptor to automate this entire journey. It acts as a digital architect that can read scientific literature, find the right drawings, turn them into 3D models, and test them against specific goals, all without human intervention.

The system works by first building a massive library of knowledge from over three thousand published papers on mechanical metamaterials. Instead of just storing text, the system learns to connect ideas, images, and formulas. It understands that a specific shape leads to a specific behavior, and it links these concepts together in a way a computer can search. When a user asks for a material with certain properties, such as a specific stiffness or a unique way of bending, the system searches this library. It finds relevant papers and extracts the structural images, even if they are cluttered with arrows, labels, or dimension lines. The system then cleans these images, removing the extra text to reveal the pure shape of the structure.

Once the image is clear, the system performs a remarkable feat: it turns the flat drawing into a three-dimensional object. It builds a digital model that looks like a tiny, intricate lattice or a porous sponge. However, a single drawing often hides parts of the structure because of the angle or shadows. To fix this, the system uses a technique called symmetry restoration. It looks at the visible parts and mirrors them across different axes, effectively guessing the hidden sides to create a complete, balanced 3D object. It then takes this object and makes small adjustments, thickening or thinning the struts that make up the structure. These small changes create a whole family of slightly different versions, each with slightly different mechanical properties.

Finally, the system puts these digital models through a rigorous test. It simulates how they would behave under pressure, calculating their stiffness and how they deform. If the user asked for a material that stretches sideways when pulled, the system checks if any of the generated models do exactly that. If the first model is close but not perfect, the system tries the next variation from its family of adjusted shapes. It repeats this process, moving through the list of recommended designs from the literature, until it finds a match that fits the user's requirements within a very tight margin of error.

The results of this automated approach are striking. When the researchers tested the system with a strict requirement for accuracy, the initial 3D models built from the drawings alone succeeded in finding a matching design only about one-third of the time. But when the system added the symmetry restoration and the small structural adjustments, the success rate jumped to over eighty percent. This means the system can now find a working design in most cases where a human would have given up or spent weeks trying to tweak the shape. It also found these matches much faster, often identifying the right structure within the first few papers it examined, rather than having to sift through dozens of possibilities.

The study demonstrates that combining artificial intelligence with the physical laws of materials can bridge the gap between what is written in a book and what can be built in a lab. The system does not just guess; it grounds its decisions in the actual images and data from published research. It links the user's request directly to a specific paper, a specific image, and a specific 3D model, creating a clear trail of how the design was found. This approach allows scientists to explore the vast landscape of possible metamaterials without being limited by their own intuition or the time it takes to build models by hand. By turning the static knowledge of the past into a dynamic tool for the future, M3Sculptor offers a new way to discover materials that can be lighter, stronger, and more adaptable than anything currently available.

Technical Summary: M3Sculptor

Problem Statement

The design of mechanical metamaterials—engineered materials with properties derived from their architecture rather than chemical composition—faces significant challenges due to the vastness of the design space. Traditional design workflows rely heavily on human intuition, requiring manual reconstruction of 2D literature illustrations into 3D CAD models, assignment of material properties, and iterative finite element analysis (FEA). This process is time-consuming and computationally expensive.

While artificial intelligence (AI) has advanced structure prediction and inverse design, existing "structure-first" approaches often treat structural variables as abstract parameters disconnected from domain knowledge. These methods frequently rediscover mechanisms already reported in literature because they cannot effectively access unstructured scientific publications. Furthermore, the transition from a structural concept to a physically verifiable simulation model requires substantial manual effort, creating a bottleneck that limits the systematic exploration of novel architectures. There is a critical need for a framework that can bridge literature knowledge retrieval, structural image understanding, 3D reconstruction, and physical verification into a single, automated workflow.

Methodology: The M3Sculptor Framework

The authors present M3Sculptor (Mechanical MetaMaterials Sculptor), a multi-agent framework designed to automate the design of mechanical metamaterials based on natural-language requirements. The system integrates Large Language Models (LLMs), multimodal models, and physics-based simulations through three interconnected layers:

1. Knowledge Layer

This layer encodes structured domain knowledge from 3,272 publications into queryable Knowledge Graphs (KGs) and multimodal databases.

Data Pipeline: The system processes PDFs using MinerU to standardize content, followed by a chat model (Qwen2.5-72B) for binary classification to filter for mechanical metamaterials.
Multimodal Alignment: A vision-language model (Qwen2.5-VL-32B) aligns figures with text captions and paragraphs.
Entity-Relation Extraction: Using a structured prompting mechanism inspired by GraphRAG, the system extracts entities (Structure, Material, Property, Mechanism, Method, Application) and their relationships, generating standardized JSON outputs.
Storage: Data is stored in a tri-layer architecture: Milvus for vector embeddings (cross-modal retrieval), Neo4j for semantic relations (traversable queries), and MongoDB for full document context.
Model Fine-tuning: The authors constructed a domain-specific dataset of 71,400 QA pairs and 24,730 multiple-choice questions (MCQs) from the literature. They fine-tuned Qwen2.5-7B and Qwen3.8-27B base models using Low-Rank Adaptation (LoRA), resulting in M3LLM-7B and M3LLM-27B.

2. Structure Layer

This layer converts 2D literature illustrations into 3D structures suitable for mechanical analysis.

Image Preprocessing: Retrieved figures are processed to remove non-structural annotations (arrows, dimension lines) using Rex-Omni and Qwen-Image-Edit.
3D Reconstruction: A diffusion transformer-based model (Hunyuan3D) generates 3D candidate meshes from the preprocessed images.
Mesh Cleaning: Trimesh software removes disconnected fragments and converts the largest connected component into a $64 \times 64 \times 64$ voxel representation.
Symmetry Restoration: To address occlusions in single-image reconstruction, the system mirrors the voxel matrix along x, y, and z axes, generating up to eight symmetric candidate structures.
Morphological Operations: The system applies erosion and dilation operators to the voxel grid to modify solid-void interfaces, generating structures with varying relative densities and local feature sizes.

3. Physics Layer

This layer evaluates candidate structures through computational homogenization.

Simulation: Periodic boundary conditions are applied to compute the effective elastic stiffness matrix ( $C^H$ ).
Property Calculation: From the compliance matrix ( $S^H$ ), the system calculates direction-dependent Young's modulus ( $E$ ), shear modulus ( $G$ ), and Poisson's ratio ( $\nu$ ).
Target Matching: The system iterates through ranked publications. For each publication, it reconstructs structures, applies transformations, and compares computed properties against user-specified targets ( $p^*$ ) within a defined tolerance ( $\delta$ ). The process stops when a match is found or the recommendation list is exhausted.

Key Results

Knowledge Graph and Model Performance

The constructed KG contains approximately 440,000 nodes and 1.7 million relationships derived from 3,272 publications.
Domain Adaptation: Fine-tuning significantly improved performance on mechanical metamaterial MCQs. M3LLM-7B achieved 76.22% accuracy (a 6.72% gain over its base model), outperforming larger general-purpose models like Qwen2.5-72B and DeepSeek-V3 on this specific domain. M3LLM-27B reached 81.04% accuracy.
General Benchmarks: While domain-specific accuracy improved, the authors noted a trade-off where MMLU (general knowledge) accuracy slightly decreased or plateaued for the fine-tuned models compared to their baselines, suggesting specialization for the metamaterial domain.

Structure Reconstruction and Property Matching

Reconstruction Success: The framework successfully reconstructed diverse architectures (e.g., zigzag beams, cuboctahedral lattices, curved surfaces) from literature images, producing meshes suitable for FEA.
Impact of Transformations: The study evaluated the contribution of structural transformations to target property matching at an absolute tolerance of $\delta = 0.01$ $δ = 0.01$ (GPa for moduli, dimensionless for Poisson's ratio):
- Initial Reconstruction Only: 34.13% success rate within five recommendation rounds.
- With Symmetry Restoration (+Sym): 53.25% success rate.
- With Morphological Operations (+Morph): 61.98% success rate.
- Expanded Set (Initial + Sym + Morph): 83.25% success rate.
Efficiency: Expanding the candidate set reduced the mean number of publications required to find a match from 4.00 (Initial) to 2.20 (Expanded), a 45% reduction.
Error Reduction: The median normalized error ( $\epsilon/\delta$ ) decreased from 2.79 (Initial) to 0.04 (Expanded), indicating a significant shift toward closer target matching.
Property Specificity: Morphological operations were particularly effective for tuning Young's and shear moduli (success rates rising to ~89-92%), while symmetry restoration was crucial for adjusting Poisson's ratio targets.

Significance and Claims

The paper claims that M3Sculptor provides a practical route for translating design experience documented in scientific literature into computationally evaluable structural designs. By integrating knowledge retrieval, image-driven 3D reconstruction, and physics-based verification, the framework addresses the disconnect between unstructured literature and automated design.

The authors emphasize that their approach moves beyond "structure-first" design by grounding the generation of candidate architectures in explicit domain knowledge and physical laws. The system demonstrates that combining symmetry restoration and morphological operations can significantly increase the probability of finding metamaterials that meet specific mechanical targets, thereby automating the exploration of vast design spaces that are otherwise inaccessible to traditional trial-and-error methods. The framework is presented as a traceable workflow where every candidate structure is linked back to its source publication, ensuring scientific reproducibility.