July 24, 2025
The GIST AI-driven framework creates defect-tolerant metamaterials with complex functionality 
Lisa Lock
scientific editor
Andrew Zinin
lead editor
Editors' notes
This article has been reviewed according to Science X's editorial process and policies. Editors have highlighted the following attributes while ensuring the content's credibility:
fact-checked
peer-reviewed publication
trusted source
proofread
Many industrial products—from car bumpers to aerospace panels and medical implants—owe their performance to lightweight, cellular materials. These hard-working synthetics are engineered to meet specific functionality goals, but too often, defects introduced during the fabrication process can lead to subpar performance or even catastrophic failure.
Now, a UC Berkeley-led team of researchers has developed a new AI-driven framework that can more efficiently design 3D truss metamaterials—a type of structure with extraordinary mechanical properties, sound absorption capabilities and tunability—while minimizing their sensitivity to defects.
In their article published in Nature Machine Intelligence, researchers demonstrate how their patent-pending modeling method, dubbed GraphMetaMat, uses deep learning techniques to bridge the gap between metamaterials design and manufacturability, paving the way for new and highly useful materials.
"Until now, most of the work done in AI and materials design has been in the theoretical and computational domain, where they give you the design that performs well under ideal conditions," said Xiaoyu (Rayne) Zheng, associate professor of materials science and engineering and the study's principal investigator.
"GraphMetaMat shows that AI can give you a realistic design tailored for a specific manufacturing method, like 3D printing, and optimized to withstand various manufacturing-related defects. It sets the stage for the automatic design of manufacturable, defect-tolerant materials with on-demand functionalities."
While advances in data-driven design and additive manufacturing have significantly accelerated the development of truss metamaterials, Zheng explained that existing inverse design approaches have inherent limitations. They can generate metamaterials with target linear properties, such as elasticity, but struggle to capture more complex nonlinear behaviors, such as energy absorption, needed for items like car bumpers and protective athletic gear.
"Design methods like topology optimization or an intuition-guided iterative approach are good at predicting simple responses," said Zheng. "But for many real-world problems, these approaches cannot efficiently design materials with the required functionality, manufacturability and tolerance to defects introduced during manufacturing."
Recently, researchers considered using graph neural networks for metamaterials design, since this has proved to be a powerful tool in drug discovery. But there was little to no training data available for designing metamaterials.
Zheng and his fellow researchers solved this problem by integrating multiple deep learning techniques—reinforcement learning, imitation learning, a surrogate model, and Monte Carlo tree search—into GraphMetaMat.
"Users can create metamaterial designs, represented as graphs, entirely from scratch based on custom inputs—such as a desired stress–strain curve or specific vibration attenuation gaps where mechanical waves are blocked at certain frequencies," said Marco Maurizi, postdoctoral researcher in the Department of Materials Science and Engineering and lead author of the study. "Our AI system then iteratively adds graph nodes and edges to define the material's geometry and topology."
Most importantly, according to Zheng, GraphMetaMat can also integrate engineering constraints into the graphs—including manufacturing and defect constraints.
"GraphMetaMat has the unique ability to account for fabrication-induced imperfections," he said. "This innovation is a game-changer because it ensures that the generated metamaterials will not fail if they develop a small defect during manufacturing."
In their proof of concept, the researchers used GraphMetaMat to design lightweight truss metamaterials optimized for energy absorption and vibration mitigation at various frequencies. For each use case, the generated metamaterial consistently outperformed traditional materials, including polymeric foams and phononic crystals.
"Based on our findings, GraphMetaMat has the potential to redefine the design paradigm," said Zheng. "This opens the door to exciting new possibilities in creating realistic, high-performance metamaterials."
More information: Marco Maurizi et al, Designing metamaterials with programmable nonlinear responses and geometric constraints in graph space, Nature Machine Intelligence (2025). DOI: 10.1038/s42256-025-01067-x
Journal information: Nature Machine Intelligence Provided by University of California – Berkeley Citation: AI-driven framework creates defect-tolerant metamaterials with complex functionality (2025, July 24) retrieved 24 July 2025 from https://techxplore.com/news/2025-07-ai-driven-framework-defect-tolerant.html This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.
Explore further
Electron microscopy technique captures nanoparticle organizations to forge new materials 2 shares
Feedback to editors
