3D-printed materials that should behave like a perfect wave filter instead rattle and wobble unpredictably because manufacturing flaws at the microscopic scale are invisible to current computer simulations. This gap between design and reality means engineers cannot reliably create metamaterials—artificially structured materials with properties not found in nature, such as being both ultralight and recoverable after compression. The problem is that traditional simulations assume perfect geometry, but real 3D-printed samples contain millions of tiny deviations that accumulate and destroy the intended performance. This project will train graph neural networks—a type of AI that learns relationships between connected elements—directly on experimental data from printed samples. The networks will learn to predict how real-world defects alter mechanical behaviour, then use that knowledge to inversely design new metamaterials: an engineer could specify a desired wave transmission spectrum, and the AI would generate a printable structure that actually delivers that response in the lab. If successful, the work would close the long-standing divide between simulation and experiment for 3D-printed metamaterials. This is fundamental science with no immediate consumer application, but it could eventually enable vibration-damping components in precision instruments, lightweight impact-absorbing structures, or acoustic filters in industrial machinery—systems where predictable mechanical behaviour matters more than novelty.
View original technical description
The emergence of additive manufacturing techniques has enabled the creation of complex 3D shapes with topological feature sizes spanning length scales from nanometres upwards. These manufacturing technologies have facilitated the creation of new materials (metamaterials) with previously unattainable properties such as light and recoverable ceramics. However, defects (deviations from the design) caused by manufacturing variabilities proliferate in topologically complex printed samples comprising millions of micro-scale elements. Traditional numerical simulations do not capture these a priori unknown defects, and thus the measured properties of fabricated metamaterials invariably deviate substantially from the designed/simulated properties. This low fidelity of the metamaterial simulation tools has left a vast portion of the metamaterial design space untapped. Leveraging recent foundational advances in machine learning and graph neural networks (GNNs), now is the ideal time for designing new additively manufactured materials with fully tailorable static and dynamic properties wherein, for example, a designer inputs a wave transmission spectrum from which a metamaterial that fully replicates the input response in an experimental setting is inversely designed and printed. Graph-based data-driven methods can address this challenge by their ability to learn from experimental data and efficiently encode the 3D material topology. The proposal will break new ground by exploiting breakthroughs in graph-based generative machine learning models to inversely generate metamaterials and thereby fuse the field of GNNs with mechanics, materials science, and additive manufacturing. This represents a fundamentally new realm of engineered material creation and discovery paradigm that will bridge the longstanding gap between simulation and experimental data of 3D printed metamaterials. The project will lay the scientific foundations for new engineering material designs and solutions.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know