Researchers Use AI to Create Super-Compressible Meta-Material
We are slowly but surely getting used to the idea of having some form of artificial intelligence in our everyday lives, from smart personal assistants to the algorithms that drive online recommendation engines. Not surprisingly, AI is also helping to accelerate the process of discovery in a wide range of fields — whether it’s finding new, life-saving pharmaceutical drugs, or innovative materials that exhibit extraordinary properties — all without relying on the usual process of experimental trial-and-error.
For researchers over at TU Delft in the Netherlands, AI was instrumental in their development of a super-compressible metamaterial out of what is typically a brittle and fragile material. With this new metamaterial, objects like bicycles or furniture could be made out of this super-compressible substance, and then squished down into something that would fit into your pocket. Take a look:
In contrast to materials that are naturally occurring, metamaterials are artificially engineered to have properties that aren’t found in nature, meaning they can do some pretty amazing things — like being “programmed” to automatically shapeshift. Though metamaterials sound high tech, more often than not the way they are developed is still pretty conventional.
“Metamaterial design has relied on extensive experimentation and a trial-and-error approach,” said the study’s co-author Miguel Bessa, who is an assistant professor in materials science and engineering at TU Delft. “We argue in favor of inverting the process by using machine learning for exploring new design possibilities while reducing experimentation to an absolute minimum.”
While at first glance this approach may seem counterintuitive to those of us who are used to the idea of employing real-world experiments to hammer out a solution, this new AI-assisted approach can make sense for companies or institutions looking for more time-saving and cost-effective ways to do research and development, allowing human experts to generate new ideas that wouldn’t have emerged otherwise.
In designing their new material, the team followed a computational data-driven approach. In particular, they focused on utilizing Bayesian machine learning, which allowed them to dynamically adapt their model on one hand in order to accommodate for different target properties and base materials, as well as a variety of scales and manufacturing methods. The AI ran through simulations that tested out the metamaterial virtually, trying out various geometries and configurations, progressively learning from the failed simulations in order to come up with something that functioned well.
“The important thing is that machine learning creates an opportunity to invert the design process by shifting from experimentally guided investigations to computationally data-driven ones, even if the computer models are missing some information,” explained Bessa. “The essential requisites are that ‘enough’ data about the problem of interest is available and that the data is sufficiently accurate.”
Using machine learning to simulate the behavior of their metamaterial, the team fabricated two designs. One design was created at the macro-scale and was geared toward maximum compressibility, thanks to its double-ringed arrangement that can twist and squeeze down to an almost-flat profile. After tweaking it virtually on their computers, the design was printed on a desktop 3D printer, via fused filament fabrication and using a polylactic acid (PLA) material, with the end result having nearly 100 percent compressibility. Similarly, the second design was a microscaled version of the metamaterial that was tailored for high strength and stiffness, exhibiting around 80% compressibility.
As one might imagine, such a data-driven approach will streamline and fine-tune the design process immensely. According to the research team, by adding AI to the mix, experts would be able to forego much of the tedious and time-consuming trial-and-error experiments that are typically done in materials science and other fields, likely revolutionizing the way we innovate and make scientific discoveries.
You can read the team’s paper here, and find the team’s code on Github.
Images: TU Delft