These synthetics are engineered to meet specific functionality goals, but defects introduced during manufacturing can lead to underperformance or, in some cases, catastrophic failure.
A research team led by the University of California, Berkeley, has developed an innovative AI-powered framework that efficiently designs 3D truss metamaterials — structures known for their exceptional mechanical strength, sound absorption, and tunability — while minimising their sensitivity to defects.
Introducing GraphMetaMat: AI for Real-World Manufacturing
Published in the Nature Machine Intelligence cover story on 22 July, the team's patent-pending modelling framework, called GraphMetaMat, uses deep learning to close the gap between theoretical design and practical manufacturability.
“Until now, most AI in materials design has been theoretical, assuming ideal conditions,” said Xiaoyu (Rayne) Zheng, associate professor of materials science and engineering and the study’s principal investigator. “GraphMetaMat shows AI can deliver realistic, defect-tolerant designs optimised for specific fabrication methods like 3D printing.”
Overcoming the Limitations of Traditional Methods
While data-driven design and additive manufacturing have accelerated metamaterial development, current inverse design techniques are often limited to predicting linear properties such as elasticity. They struggle with nonlinear behaviours like energy absorption — critical for applications such as impact-resistant gear.
“Traditional approaches, including topology optimisation or iterative intuition-guided methods, fall short when designing for complex functionalities and manufacturing realities,” Zheng noted.
Leveraging Deep Learning for Complex Functionality
To address the challenge, Zheng's team turned to graph neural networks — a technique gaining traction in drug discovery — but adapted them for metamaterials. Due to the lack of existing training data, the researchers built GraphMetaMat using a fusion of deep learning techniques, including reinforcement learning, imitation learning, surrogate modelling, and Monte Carlo tree search.
“Users can generate new metamaterial designs as graphs, based on inputs like desired stress–strain curves or frequency-specific vibration mitigation,” said Marco Maurizi, postdoctoral researcher and lead author. “The AI then builds the material’s geometry and topology node by node.”
Designed for Real-World Defects
What sets GraphMetaMat apart is its ability to incorporate real-world engineering constraints directly into the design — including tolerance to manufacturing-induced defects.
“This is a game-changer,” said Zheng. “GraphMetaMat ensures the resulting materials maintain functionality even with imperfections from the fabrication process.”
Outperforming Traditional Materials
In a proof-of-concept, the researchers applied GraphMetaMat to develop lightweight truss metamaterials aimed at energy absorption and vibration mitigation across multiple frequencies. In every test, the AI-designed materials outperformed traditional options like polymer foams and phononic crystals.
“GraphMetaMat has the potential to redefine the way we approach material design,” Zheng said. “It unlocks exciting opportunities for creating high-performance metamaterials tailored for real-world applications.”
Collaborative Effort and Funding
This groundbreaking research was conducted in collaboration with UCLA (led by Wei Wang and Yizhou Sun) and Penn State University (led by Yun Jing). Co-lead authors include Derek Xu (UCLA) and Yu-Tong Wang (Penn State University). The project received primary funding from the NSF’s Designing Materials to Revolutionise and Engineer our Future (DMREF) initiative.
